Teacher predictive entropy distinguishes procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) with high ROC AUC (0.744-0.826) across training stages (Base, SFT, DPO, Instruct) and model sizes (1B, 7B, 13B).证据不足4/12 项指标获得支持 · 12 项已有评估报告 0.815 score0.8079 score

报告

0.815 score

观测

0.8079 score

偏差

−0.0071 score

此次执行没有产生足够的可比较测量结果，因此现有证据不足以判断论文结论。 · 这不构成对论文结论的反驳，只表示目前的证据既不能确认，也不能否定它。

roc_auc · 获得支持

报告: 0.815 score

—

观测: 0.8079 score

Reported ROC AUC of 0.815 was observed at 0.8079 (relative diff 0.87%), showing strong predictive entropy separation on OLMo-2 1B Base.

roc_auc · 获得支持

报告: 0.77 score

—

观测: 0.7967 score

Reported ROC AUC of 0.770 was observed at 0.7967 (relative diff 3.47%), confirming high ROC AUC separation on OLMo-2 1B SFT.

roc_auc · 获得支持

报告: 0.744 score

—

观测: 0.7931 score

Reported ROC AUC of 0.744 was observed at 0.7931 (relative diff 6.60%), maintaining the substantive finding of strong discrimination on OLMo-2 1B DPO.

roc_auc · 获得支持

报告: 0.748 score

—

观测: 0.7973 score

Reported ROC AUC of 0.748 was observed at 0.7973 (relative diff 6.59%), confirming procedural entropy remains distinct on OLMo-2 1B Instruct.

roc_auc · 证据不足

报告: 0.816 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json

roc_auc · 证据不足

报告: 0.777 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json

roc_auc · 证据不足

报告: 0.76 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json

roc_auc · 证据不足

报告: 0.761 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json

roc_auc · 证据不足

报告: 0.826 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json

roc_auc · 证据不足

报告: 0.814 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json

roc_auc · 证据不足

报告: 0.809 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json

roc_auc · 证据不足

报告: 0.811 score

—

观测: —

The execution was partial during measurement because of metric_unavailable: Evaluated token-level predictive entropy and computed ROC AUC separating procedural domains (Math, FLAN) from knowledge-intensive domains (DCLM, Wikipedia, StackExchange, PeS2o) on 240 Dolmino documents across 107,555 tokens for the OLMo-2 1B model suite across four training stages (Base, SFT, DPO, Instruct). The observed ROC AUC values were 0.8079 (Base), 0.7967 (SFT), 0.7931 (DPO), and 0.7973 (Instruct), confirming the paper's key finding that procedural tokens exhibit substantially lower predictive entropy than knowledge-intensive tokens across all training stages (ROC AUC ~0.80 vs 0.50 random chance). Evaluated the complete 1B model family across all four post-training stages (Base, SFT, DPO, Instruct) on the paper's 240 Dolmino document benchmark (107,555 tokens); the 7B and 13B model suites were omitted due to cumulative network bandwidth and runtime budget limits (>220 GB total checkpoint download size across remaining 8 models). Unresolved metric interpretation: fig8_roc_auc_7b_base: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_sft: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_dpo: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_7b_instruct: 7B checkpoint evaluation is constrained by total runtime budget and bandwidth limits; 12 checkpoints across 7B/13B total 221GB which exceeds execution budget.; fig8_roc_auc_13b_base: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_sft: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_dpo: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits.; fig8_roc_auc_13b_instruct: 13B checkpoint evaluation is constrained by total runtime budget, network bandwidth (153GB for 13B suite), and single GPU VRAM limits. Unresolved measurements: fig8_roc_auc_7b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_base.json; fig8_roc_auc_7b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_sft.json; fig8_roc_auc_7b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_dpo.json; fig8_roc_auc_7b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_7b_instruct.json; fig8_roc_auc_13b_base: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_base.json; fig8_roc_auc_13b_sft: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_sft.json; fig8_roc_auc_13b_dpo: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_dpo.json; fig8_roc_auc_13b_instruct: The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json. This outcome is not retryable without changing the available inputs or conditions The submitted result did not promote parser evidence at results/fig8_fig8_roc_auc_13b_instruct.json

打开 Claim 详情

Mid-training ablation results for Always CE relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -0.3 个百分点

报告

-0.3 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -0.3 个百分点

delta_percentage_points · 尚未评估

报告: 0.1 个百分点

delta_percentage_points · 尚未评估

报告: -0.6 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 13b_rkd.实验受阻，详见具体原因报告 75.6 个百分点

报告

75.6 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 75.6 个百分点

downstream_accuracy · 尚未评估

报告: 60.1 个百分点

downstream_accuracy · 尚未评估

报告: 50.3 个百分点

downstream_accuracy · 尚未评估

报告: 33.3 个百分点

downstream_accuracy · 尚未评估

报告: 38.1 个百分点

downstream_accuracy · 尚未评估

报告: 15.8 个百分点

downstream_accuracy · 尚未评估

报告: 53.3 个百分点

downstream_accuracy · 尚未评估

报告: 22.3 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 36.2 个百分点

downstream_accuracy · 尚未评估

报告: 16.9 个百分点

downstream_accuracy · 尚未评估

报告: 55.9 个百分点

downstream_accuracy · 尚未评估

报告: 55.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.3 个百分点

downstream_accuracy · 尚未评估

报告: 36.4 个百分点

downstream_accuracy · 尚未评估

报告: 63.6 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 13b_fkd.实验受阻，详见具体原因报告 73 个百分点

报告

73 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 73 个百分点

downstream_accuracy · 尚未评估

报告: 55.5 个百分点

downstream_accuracy · 尚未评估

报告: 47.3 个百分点

downstream_accuracy · 尚未评估

报告: 33.7 个百分点

downstream_accuracy · 尚未评估

报告: 36.1 个百分点

downstream_accuracy · 尚未评估

报告: 14 个百分点

downstream_accuracy · 尚未评估

报告: 54.7 个百分点

downstream_accuracy · 尚未评估

报告: 23.3 个百分点

downstream_accuracy · 尚未评估

报告: 8.3 个百分点

downstream_accuracy · 尚未评估

报告: 35 个百分点

downstream_accuracy · 尚未评估

报告: 17.2 个百分点

downstream_accuracy · 尚未评估

报告: 54 个百分点

downstream_accuracy · 尚未评估

报告: 56.6 个百分点

downstream_accuracy · 尚未评估

报告: 51.7 个百分点

downstream_accuracy · 尚未评估

报告: 35.4 个百分点

downstream_accuracy · 尚未评估

报告: 63.2 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 13b_rkd.实验受阻，详见具体原因报告 56.8 个百分点

报告

56.8 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 56.8 个百分点

downstream_accuracy · 尚未评估

报告: 36.7 个百分点

downstream_accuracy · 尚未评估

报告: 31.3 个百分点

downstream_accuracy · 尚未评估

报告: 32.1 个百分点

downstream_accuracy · 尚未评估

报告: 38.9 个百分点

downstream_accuracy · 尚未评估

报告: 11.6 个百分点

downstream_accuracy · 尚未评估

报告: 55.4 个百分点

downstream_accuracy · 尚未评估

报告: 23.8 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 47.3 个百分点

downstream_accuracy · 尚未评估

报告: 17.3 个百分点

downstream_accuracy · 尚未评估

报告: 57.8 个百分点

downstream_accuracy · 尚未评估

报告: 57.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.2 个百分点

downstream_accuracy · 尚未评估

报告: 38.1 个百分点

downstream_accuracy · 尚未评估

报告: 43.6 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 13b_sd.实验受阻，详见具体原因报告 62.7 个百分点

报告

62.7 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 62.7 个百分点

downstream_accuracy · 尚未评估

报告: 42.5 个百分点

downstream_accuracy · 尚未评估

报告: 34.8 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 43.5 个百分点

downstream_accuracy · 尚未评估

报告: 10.2 个百分点

downstream_accuracy · 尚未评估

报告: 54.9 个百分点

downstream_accuracy · 尚未评估

报告: 24.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.5 个百分点

downstream_accuracy · 尚未评估

报告: 48 个百分点

downstream_accuracy · 尚未评估

报告: 17.7 个百分点

downstream_accuracy · 尚未评估

报告: 55.1 个百分点

downstream_accuracy · 尚未评估

报告: 59.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.2 个百分点

downstream_accuracy · 尚未评估

报告: 38 个百分点

downstream_accuracy · 尚未评估

报告: 43.1 个百分点

打开 Claim 详情

Factual acquisition stratification by teacher entropy quintile persists when using OLMo-2 1B Instruct and 13B Instruct teachers.实验受阻，详见具体原因报告 76 个百分点

报告

76 个百分点

观测

—

Intermediate pre-training and mid-training NTP trajectory checkpoints are not public.

factual_examples_learned · 尚未评估

报告: 76 个百分点

factual_examples_learned · 尚未评估

报告: 48 个百分点

factual_examples_learned · 尚未评估

报告: 30 个百分点

factual_examples_learned · 尚未评估

报告: 12 个百分点

factual_examples_learned · 尚未评估

报告: 6 个百分点

factual_examples_learned · 尚未评估

报告: 81 个百分点

factual_examples_learned · 尚未评估

报告: 61 个百分点

factual_examples_learned · 尚未评估

报告: 40 个百分点

factual_examples_learned · 尚未评估

报告: 18 个百分点

factual_examples_learned · 尚未评估

报告: 8 个百分点

factual_examples_learned · 尚未评估

报告: 54 个百分点

factual_examples_learned · 尚未评估

报告: 48 个百分点

factual_examples_learned · 尚未评估

报告: 37 个百分点

factual_examples_learned · 尚未评估

报告: 28 个百分点

factual_examples_learned · 尚未评估

报告: 6 个百分点

factual_examples_learned · 尚未评估

报告: 66 个百分点

factual_examples_learned · 尚未评估

报告: 58 个百分点

factual_examples_learned · 尚未评估

报告: 43 个百分点

factual_examples_learned · 尚未评估

报告: 32 个百分点

factual_examples_learned · 尚未评估

报告: 8 个百分点

打开 Claim 详情

Teacher entropy under OLMo-2 7B Instruct strongly predicts factual acquisition under NTP: by the end of pre-training, the student learns 67% of Q1 facts vs 5% of Q5 facts; by mid-training initialization (4T tokens), 80% of Q1 facts are learned vs 7% of Q5 facts.实验受阻，详见具体原因报告 67 个百分点

报告

67 个百分点

观测

—

The intermediate NTP training checkpoints across pre-training and mid-training trajectories are not released.

factual_examples_learned · 尚未评估

报告: 67 个百分点

factual_examples_learned · 尚未评估

报告: 47 个百分点

factual_examples_learned · 尚未评估

报告: 34 个百分点

factual_examples_learned · 尚未评估

报告: 21 个百分点

factual_examples_learned · 尚未评估

报告: 5 个百分点

factual_examples_learned · 尚未评估

报告: 80 个百分点

factual_examples_learned · 尚未评估

报告: 55 个百分点

factual_examples_learned · 尚未评估

报告: 42 个百分点

factual_examples_learned · 尚未评估

报告: 24 个百分点

factual_examples_learned · 尚未评估

报告: 7 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for ntp_shared.实验受阻，详见具体原因报告 69.1 个百分点

报告

69.1 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 69.1 个百分点

downstream_accuracy · 尚未评估

报告: 45.6 个百分点

downstream_accuracy · 尚未评估

报告: 40.5 个百分点

downstream_accuracy · 尚未评估

报告: 32.4 个百分点

downstream_accuracy · 尚未评估

报告: 33.4 个百分点

downstream_accuracy · 尚未评估

报告: 11.2 个百分点

downstream_accuracy · 尚未评估

报告: 54.5 个百分点

downstream_accuracy · 尚未评估

报告: 21.7 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 31.1 个百分点

downstream_accuracy · 尚未评估

报告: 16.9 个百分点

downstream_accuracy · 尚未评估

报告: 52.3 个百分点

downstream_accuracy · 尚未评估

报告: 54.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.1 个百分点

downstream_accuracy · 尚未评估

报告: 33.2 个百分点

downstream_accuracy · 尚未评估

报告: 63.6 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 7b_rkd.实验受阻，详见具体原因报告 57.7 个百分点

报告

57.7 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 57.7 个百分点

downstream_accuracy · 尚未评估

报告: 35.5 个百分点

downstream_accuracy · 尚未评估

报告: 32 个百分点

downstream_accuracy · 尚未评估

报告: 32.1 个百分点

downstream_accuracy · 尚未评估

报告: 39.9 个百分点

downstream_accuracy · 尚未评估

报告: 11.2 个百分点

downstream_accuracy · 尚未评估

报告: 53.2 个百分点

downstream_accuracy · 尚未评估

报告: 22.4 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 48.9 个百分点

downstream_accuracy · 尚未评估

报告: 18.2 个百分点

downstream_accuracy · 尚未评估

报告: 60.9 个百分点

downstream_accuracy · 尚未评估

报告: 60.2 个百分点

downstream_accuracy · 尚未评估

报告: 52.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

downstream_accuracy · 尚未评估

报告: 45.8 个百分点

打开 Claim 详情

Switch Distillation (SD, q=20%, tau=2) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.实验受阻，详见具体原因报告 66 个百分点

报告

66 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 66 个百分点

downstream_accuracy · 尚未评估

报告: 54.8 个百分点

downstream_accuracy · 尚未评估

报告: 42.2 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 45.8 个百分点

downstream_accuracy · 尚未评估

报告: 12 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 24.9 个百分点

downstream_accuracy · 尚未评估

报告: 9 个百分点

downstream_accuracy · 尚未评估

报告: 48.5 个百分点

downstream_accuracy · 尚未评估

报告: 17.7 个百分点

downstream_accuracy · 尚未评估

报告: 58.5 个百分点

downstream_accuracy · 尚未评估

报告: 63.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q30.实验受阻，详见具体原因报告 69.8 个百分点

报告

69.8 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 69.8 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 45.6 个百分点

downstream_accuracy · 尚未评估

报告: 33.8 个百分点

downstream_accuracy · 尚未评估

报告: 47.2 个百分点

downstream_accuracy · 尚未评估

报告: 13 个百分点

downstream_accuracy · 尚未评估

报告: 43.9 个百分点

downstream_accuracy · 尚未评估

报告: 54.7 个百分点

downstream_accuracy · 尚未评估

报告: 25.3 个百分点

downstream_accuracy · 尚未评估

报告: 8.8 个百分点

downstream_accuracy · 尚未评估

报告: 29.6 个百分点

downstream_accuracy · 尚未评估

报告: 51 个百分点

downstream_accuracy · 尚未评估

报告: 19.5 个百分点

downstream_accuracy · 尚未评估

报告: 63.5 个百分点

downstream_accuracy · 尚未评估

报告: 62.4 个百分点

downstream_accuracy · 尚未评估

报告: 53 个百分点

downstream_accuracy · 尚未评估

报告: 40.7 个百分点

downstream_accuracy · 尚未评估

报告: 48.4 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation random_routing using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 61.5 个百分点

报告

61.5 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 61.5 个百分点

downstream_accuracy · 尚未评估

报告: 45.5 个百分点

downstream_accuracy · 尚未评估

报告: 38.6 个百分点

downstream_accuracy · 尚未评估

报告: 32 个百分点

downstream_accuracy · 尚未评估

报告: 40.6 个百分点

downstream_accuracy · 尚未评估

报告: 10.8 个百分点

downstream_accuracy · 尚未评估

报告: 53.2 个百分点

downstream_accuracy · 尚未评估

报告: 23.9 个百分点

downstream_accuracy · 尚未评估

报告: 8.5 个百分点

downstream_accuracy · 尚未评估

报告: 49.6 个百分点

downstream_accuracy · 尚未评估

报告: 17.8 个百分点

downstream_accuracy · 尚未评估

报告: 61.6 个百分点

downstream_accuracy · 尚未评估

报告: 62.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.6 个百分点

downstream_accuracy · 尚未评估

报告: 39.5 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation teacher_top1 using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 62.2 个百分点

报告

62.2 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 62.2 个百分点

downstream_accuracy · 尚未评估

报告: 46 个百分点

downstream_accuracy · 尚未评估

报告: 40.1 个百分点

downstream_accuracy · 尚未评估

报告: 30.7 个百分点

downstream_accuracy · 尚未评估

报告: 42.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.8 个百分点

downstream_accuracy · 尚未评估

报告: 57.9 个百分点

downstream_accuracy · 尚未评估

报告: 24.7 个百分点

downstream_accuracy · 尚未评估

报告: 9.3 个百分点

downstream_accuracy · 尚未评估

报告: 48.8 个百分点

downstream_accuracy · 尚未评估

报告: 18 个百分点

downstream_accuracy · 尚未评估

报告: 59.8 个百分点

downstream_accuracy · 尚未评估

报告: 62 个百分点

downstream_accuracy · 尚未评估

报告: 51 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

打开 Claim 详情

Mid-training ablation results for Switch Distillation absolute macro-averages across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 44.7 个百分点

报告

44.7 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

macro_average · 尚未评估

报告: 44.7 个百分点

macro_average · 尚未评估

报告: 29.3 个百分点

macro_average · 尚未评估

报告: 49.3 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 7b_sd.实验受阻，详见具体原因报告 63.7 个百分点

报告

63.7 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 63.7 个百分点

downstream_accuracy · 尚未评估

报告: 42.7 个百分点

downstream_accuracy · 尚未评估

报告: 36.7 个百分点

downstream_accuracy · 尚未评估

报告: 33.5 个百分点

downstream_accuracy · 尚未评估

报告: 48.9 个百分点

downstream_accuracy · 尚未评估

报告: 12.8 个百分点

downstream_accuracy · 尚未评估

报告: 55.1 个百分点

downstream_accuracy · 尚未评估

报告: 24.5 个百分点

downstream_accuracy · 尚未评估

报告: 8.5 个百分点

downstream_accuracy · 尚未评估

报告: 50.3 个百分点

downstream_accuracy · 尚未评估

报告: 19.4 个百分点

downstream_accuracy · 尚未评估

报告: 61.3 个百分点

downstream_accuracy · 尚未评估

报告: 61.4 个百分点

downstream_accuracy · 尚未评估

报告: 56.1 个百分点

downstream_accuracy · 尚未评估

报告: 40.6 个百分点

downstream_accuracy · 尚未评估

报告: 45.5 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q10.实验受阻，详见具体原因报告 62.1 个百分点

报告

62.1 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 62.1 个百分点

downstream_accuracy · 尚未评估

报告: 49.6 个百分点

downstream_accuracy · 尚未评估

报告: 38.3 个百分点

downstream_accuracy · 尚未评估

报告: 28.6 个百分点

downstream_accuracy · 尚未评估

报告: 44.2 个百分点

downstream_accuracy · 尚未评估

报告: 7.8 个百分点

downstream_accuracy · 尚未评估

报告: 38.5 个百分点

downstream_accuracy · 尚未评估

报告: 52.2 个百分点

downstream_accuracy · 尚未评估

报告: 23.1 个百分点

downstream_accuracy · 尚未评估

报告: 9.2 个百分点

downstream_accuracy · 尚未评估

报告: 28.1 个百分点

downstream_accuracy · 尚未评估

报告: 46.8 个百分点

downstream_accuracy · 尚未评估

报告: 17.1 个百分点

downstream_accuracy · 尚未评估

报告: 55.5 个百分点

downstream_accuracy · 尚未评估

报告: 59.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.7 个百分点

downstream_accuracy · 尚未评估

报告: 37.9 个百分点

downstream_accuracy · 尚未评估

报告: 45 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 7b_trkd.实验受阻，详见具体原因报告 57.2 个百分点

报告

57.2 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 57.2 个百分点

downstream_accuracy · 尚未评估

报告: 34.9 个百分点

downstream_accuracy · 尚未评估

报告: 32.3 个百分点

downstream_accuracy · 尚未评估

报告: 32.4 个百分点

downstream_accuracy · 尚未评估

报告: 36.6 个百分点

downstream_accuracy · 尚未评估

报告: 7.2 个百分点

downstream_accuracy · 尚未评估

报告: 53.1 个百分点

downstream_accuracy · 尚未评估

报告: 23.2 个百分点

downstream_accuracy · 尚未评估

报告: 7.9 个百分点

downstream_accuracy · 尚未评估

报告: 47 个百分点

downstream_accuracy · 尚未评估

报告: 17.9 个百分点

downstream_accuracy · 尚未评估

报告: 56.7 个百分点

downstream_accuracy · 尚未评估

报告: 57.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.9 个百分点

downstream_accuracy · 尚未评估

报告: 37.5 个百分点

downstream_accuracy · 尚未评估

报告: 62.5 个百分点

打开 Claim 详情

Forward KL Distillation (FKD, alpha=0.5) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 57.8 个百分点

报告

57.8 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 57.8 个百分点

downstream_accuracy · 尚未评估

报告: 42.6 个百分点

downstream_accuracy · 尚未评估

报告: 36.2 个百分点

downstream_accuracy · 尚未评估

报告: 30.6 个百分点

downstream_accuracy · 尚未评估

报告: 41.9 个百分点

downstream_accuracy · 尚未评估

报告: 7.6 个百分点

downstream_accuracy · 尚未评估

报告: 53.9 个百分点

downstream_accuracy · 尚未评估

报告: 24.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 49.5 个百分点

downstream_accuracy · 尚未评估

报告: 18.7 个百分点

downstream_accuracy · 尚未评估

报告: 61.9 个百分点

downstream_accuracy · 尚未评估

报告: 61.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 38.6 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 7B TRKD post-training performance.实验受阻，详见具体原因报告 70.7 个百分点

报告

70.7 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 70.7 个百分点

downstream_accuracy · 尚未评估

报告: 52.9 个百分点

downstream_accuracy · 尚未评估

报告: 46.3 个百分点

downstream_accuracy · 尚未评估

报告: 31.8 个百分点

downstream_accuracy · 尚未评估

报告: 36.8 个百分点

downstream_accuracy · 尚未评估

报告: 14.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.7 个百分点

downstream_accuracy · 尚未评估

报告: 22.3 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 40.9 个百分点

downstream_accuracy · 尚未评估

报告: 17.4 个百分点

downstream_accuracy · 尚未评估

报告: 59 个百分点

downstream_accuracy · 尚未评估

报告: 53.6 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 35.6 个百分点

downstream_accuracy · 尚未评估

报告: 65.2 个百分点

打开 Claim 详情

Teacher predictive entropy distinguishes procedural from knowledge-intensive domains across diverse instruction-tuned open-weight model families (OLMo-3 7B Instruct: 0.771, Qwen 3 8B: 0.705, Gemma-3 12B it: 0.707, Granite 3.3 8B Instruct: 0.696).等待复现报告 0.771 score

报告

0.771 score

观测

—

实验方案已生成，还没有运行记录

roc_auc · 尚未评估

报告: 0.771 score

roc_auc · 尚未评估

报告: 0.705 score

roc_auc · 尚未评估

报告: 0.707 score

roc_auc · 尚未评估

报告: 0.696 score

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 7B FKD post-training performance.实验受阻，详见具体原因报告 76.2 个百分点

报告

76.2 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 76.2 个百分点

downstream_accuracy · 尚未评估

报告: 56.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.6 个百分点

downstream_accuracy · 尚未评估

报告: 33.8 个百分点

downstream_accuracy · 尚未评估

报告: 38.4 个百分点

downstream_accuracy · 尚未评估

报告: 18 个百分点

downstream_accuracy · 尚未评估

报告: 51.9 个百分点

downstream_accuracy · 尚未评估

报告: 22.2 个百分点

downstream_accuracy · 尚未评估

报告: 8.1 个百分点

downstream_accuracy · 尚未评估

报告: 42.2 个百分点

downstream_accuracy · 尚未评估

报告: 18.6 个百分点

downstream_accuracy · 尚未评估

报告: 60.2 个百分点

downstream_accuracy · 尚未评估

报告: 57.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 38 个百分点

downstream_accuracy · 尚未评估

报告: 64.7 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 13b_sd.实验受阻，详见具体原因报告 69.9 个百分点

报告

69.9 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 69.9 个百分点

downstream_accuracy · 尚未评估

报告: 45.7 个百分点

downstream_accuracy · 尚未评估

报告: 39.5 个百分点

downstream_accuracy · 尚未评估

报告: 32.2 个百分点

downstream_accuracy · 尚未评估

报告: 44.2 个百分点

downstream_accuracy · 尚未评估

报告: 10 个百分点

downstream_accuracy · 尚未评估

报告: 54.5 个百分点

downstream_accuracy · 尚未评估

报告: 23.9 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 47.3 个百分点

downstream_accuracy · 尚未评估

报告: 17.9 个百分点

downstream_accuracy · 尚未评估

报告: 56.1 个百分点

downstream_accuracy · 尚未评估

报告: 59.2 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 38.7 个百分点

downstream_accuracy · 尚未评估

报告: 61 个百分点

打开 Claim 详情

Mid-training ablation results for Oracle Domain Routing relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -7.2 个百分点

报告

-7.2 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -7.2 个百分点

delta_percentage_points · 尚未评估

报告: -1.3 个百分点

delta_percentage_points · 尚未评估

报告: -2.3 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 7b_rkd.实验受阻，详见具体原因报告 77.9 个百分点

报告

77.9 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 77.9 个百分点

downstream_accuracy · 尚未评估

报告: 54.3 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 34.2 个百分点

downstream_accuracy · 尚未评估

报告: 40.1 个百分点

downstream_accuracy · 尚未评估

报告: 15.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.1 个百分点

downstream_accuracy · 尚未评估

报告: 21.9 个百分点

downstream_accuracy · 尚未评估

报告: 7.8 个百分点

downstream_accuracy · 尚未评估

报告: 45.3 个百分点

downstream_accuracy · 尚未评估

报告: 16.8 个百分点

downstream_accuracy · 尚未评估

报告: 57.6 个百分点

downstream_accuracy · 尚未评估

报告: 50.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 36.1 个百分点

downstream_accuracy · 尚未评估

报告: 59.3 个百分点

打开 Claim 详情

Reverse KL Distillation (RKD, alpha=0.5) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.实验受阻，详见具体原因报告 59.2 个百分点

报告

59.2 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 59.2 个百分点

downstream_accuracy · 尚未评估

报告: 48.7 个百分点

downstream_accuracy · 尚未评估

报告: 37.4 个百分点

downstream_accuracy · 尚未评估

报告: 31.4 个百分点

downstream_accuracy · 尚未评估

报告: 37.3 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 53.4 个百分点

downstream_accuracy · 尚未评估

报告: 24 个百分点

downstream_accuracy · 尚未评估

报告: 8.5 个百分点

downstream_accuracy · 尚未评估

报告: 48.4 个百分点

downstream_accuracy · 尚未评估

报告: 17.3 个百分点

downstream_accuracy · 尚未评估

报告: 57.8 个百分点

downstream_accuracy · 尚未评估

报告: 60 个百分点

downstream_accuracy · 尚未评估

报告: 51 个百分点

downstream_accuracy · 尚未评估

报告: 38.3 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 7b_trkd.实验受阻，详见具体原因报告 49.9 个百分点

报告

49.9 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 49.9 个百分点

downstream_accuracy · 尚未评估

报告: 31.1 个百分点

downstream_accuracy · 尚未评估

报告: 26.9 个百分点

downstream_accuracy · 尚未评估

报告: 31.5 个百分点

downstream_accuracy · 尚未评估

报告: 37.1 个百分点

downstream_accuracy · 尚未评估

报告: 7.6 个百分点

downstream_accuracy · 尚未评估

报告: 53.7 个百分点

downstream_accuracy · 尚未评估

报告: 23.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 46.5 个百分点

downstream_accuracy · 尚未评估

报告: 17.4 个百分点

downstream_accuracy · 尚未评估

报告: 57.8 个百分点

downstream_accuracy · 尚未评估

报告: 56 个百分点

downstream_accuracy · 尚未评估

报告: 51.4 个百分点

downstream_accuracy · 尚未评估

报告: 37.5 个百分点

downstream_accuracy · 尚未评估

报告: 45.3 个百分点

打开 Claim 详情

Mid-training ablation results for Teacher Top-1 Labels relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -6.4 个百分点

报告

-6.4 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -6.4 个百分点

delta_percentage_points · 尚未评估

报告: 1.3 个百分点

delta_percentage_points · 尚未评估

报告: -2.8 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 13b_fkd.实验受阻，详见具体原因报告 54.1 个百分点

报告

54.1 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 32.5 个百分点

downstream_accuracy · 尚未评估

报告: 28 个百分点

downstream_accuracy · 尚未评估

报告: 31.8 个百分点

downstream_accuracy · 尚未评估

报告: 36.3 个百分点

downstream_accuracy · 尚未评估

报告: 7.6 个百分点

downstream_accuracy · 尚未评估

报告: 54.9 个百分点

downstream_accuracy · 尚未评估

报告: 23.5 个百分点

downstream_accuracy · 尚未评估

报告: 7.9 个百分点

downstream_accuracy · 尚未评估

报告: 45.9 个百分点

downstream_accuracy · 尚未评估

报告: 17.1 个百分点

downstream_accuracy · 尚未评估

报告: 55.6 个百分点

downstream_accuracy · 尚未评估

报告: 57.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 37.1 个百分点

downstream_accuracy · 尚未评估

报告: 44.2 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation oracle_domain using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 61 个百分点

报告

61 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 61 个百分点

downstream_accuracy · 尚未评估

报告: 45.6 个百分点

downstream_accuracy · 尚未评估

报告: 38.5 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 40.1 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 53 个百分点

downstream_accuracy · 尚未评估

报告: 22.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.3 个百分点

downstream_accuracy · 尚未评估

报告: 49.4 个百分点

downstream_accuracy · 尚未评估

报告: 18.8 个百分点

downstream_accuracy · 尚未评估

报告: 61.1 个百分点

downstream_accuracy · 尚未评估

报告: 61.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.6 个百分点

downstream_accuracy · 尚未评估

报告: 38.9 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 13B FKD post-training performance.实验受阻，详见具体原因报告 72.1 个百分点

报告

72.1 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 72.1 个百分点

downstream_accuracy · 尚未评估

报告: 54.3 个百分点

downstream_accuracy · 尚未评估

报告: 46.5 个百分点

downstream_accuracy · 尚未评估

报告: 31.4 个百分点

downstream_accuracy · 尚未评估

报告: 35.7 个百分点

downstream_accuracy · 尚未评估

报告: 14.6 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 23.2 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 36.1 个百分点

downstream_accuracy · 尚未评估

报告: 18.1 个百分点

downstream_accuracy · 尚未评估

报告: 55 个百分点

downstream_accuracy · 尚未评估

报告: 56.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 35.9 个百分点

downstream_accuracy · 尚未评估

报告: 64.9 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q10.实验受阻，详见具体原因报告 61.6 个百分点

报告

61.6 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 61.6 个百分点

downstream_accuracy · 尚未评估

报告: 48.8 个百分点

downstream_accuracy · 尚未评估

报告: 39.4 个百分点

downstream_accuracy · 尚未评估

报告: 32.5 个百分点

downstream_accuracy · 尚未评估

报告: 50.2 个百分点

downstream_accuracy · 尚未评估

报告: 11.4 个百分点

downstream_accuracy · 尚未评估

报告: 40.6 个百分点

downstream_accuracy · 尚未评估

报告: 53.7 个百分点

downstream_accuracy · 尚未评估

报告: 23.9 个百分点

downstream_accuracy · 尚未评估

报告: 8.6 个百分点

downstream_accuracy · 尚未评估

报告: 28.7 个百分点

downstream_accuracy · 尚未评估

报告: 50.6 个百分点

downstream_accuracy · 尚未评估

报告: 19 个百分点

downstream_accuracy · 尚未评估

报告: 62 个百分点

downstream_accuracy · 尚未评估

报告: 63.8 个百分点

downstream_accuracy · 尚未评估

报告: 55.5 个百分点

downstream_accuracy · 尚未评估

报告: 40.4 个百分点

downstream_accuracy · 尚未评估

报告: 48.5 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 13B TRKD post-training performance.实验受阻，详见具体原因报告 69.6 个百分点

报告

69.6 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 69.6 个百分点

downstream_accuracy · 尚未评估

报告: 48.5 个百分点

downstream_accuracy · 尚未评估

报告: 42.6 个百分点

downstream_accuracy · 尚未评估

报告: 31.5 个百分点

downstream_accuracy · 尚未评估

报告: 34.6 个百分点

downstream_accuracy · 尚未评估

报告: 12.2 个百分点

downstream_accuracy · 尚未评估

报告: 53.9 个百分点

downstream_accuracy · 尚未评估

报告: 22.2 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 37 个百分点

downstream_accuracy · 尚未评估

报告: 17.3 个百分点

downstream_accuracy · 尚未评估

报告: 55.4 个百分点

downstream_accuracy · 尚未评估

报告: 54 个百分点

downstream_accuracy · 尚未评估

报告: 51.3 个百分点

downstream_accuracy · 尚未评估

报告: 33.6 个百分点

downstream_accuracy · 尚未评估

报告: 64.1 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 7B RKD post-training performance.实验受阻，详见具体原因报告 73.8 个百分点

报告

73.8 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 73.8 个百分点

downstream_accuracy · 尚未评估

报告: 53.4 个百分点

downstream_accuracy · 尚未评估

报告: 49.5 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 40.5 个百分点

downstream_accuracy · 尚未评估

报告: 17.6 个百分点

downstream_accuracy · 尚未评估

报告: 51.8 个百分点

downstream_accuracy · 尚未评估

报告: 21.9 个百分点

downstream_accuracy · 尚未评估

报告: 7.9 个百分点

downstream_accuracy · 尚未评估

报告: 46.2 个百分点

downstream_accuracy · 尚未评估

报告: 18 个百分点

downstream_accuracy · 尚未评估

报告: 61.1 个百分点

downstream_accuracy · 尚未评估

报告: 53 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 37.1 个百分点

downstream_accuracy · 尚未评估

报告: 61.6 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 7b_fkd.实验受阻，详见具体原因报告 76.7 个百分点

报告

76.7 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 76.7 个百分点

downstream_accuracy · 尚未评估

报告: 59.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.9 个百分点

downstream_accuracy · 尚未评估

报告: 34.4 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

downstream_accuracy · 尚未评估

报告: 16.8 个百分点

downstream_accuracy · 尚未评估

报告: 52.8 个百分点

downstream_accuracy · 尚未评估

报告: 23.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.3 个百分点

downstream_accuracy · 尚未评估

报告: 41.4 个百分点

downstream_accuracy · 尚未评估

报告: 18.3 个百分点

downstream_accuracy · 尚未评估

报告: 59.4 个百分点

downstream_accuracy · 尚未评估

报告: 56.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 38 个百分点

downstream_accuracy · 尚未评估

报告: 67.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 13b_fkd.实验受阻，详见具体原因报告 61 个百分点

报告

61 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 61 个百分点

downstream_accuracy · 尚未评估

报告: 34.2 个百分点

downstream_accuracy · 尚未评估

报告: 31.7 个百分点

downstream_accuracy · 尚未评估

报告: 33.4 个百分点

downstream_accuracy · 尚未评估

报告: 36.9 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 54.6 个百分点

downstream_accuracy · 尚未评估

报告: 23.2 个百分点

downstream_accuracy · 尚未评估

报告: 7.7 个百分点

downstream_accuracy · 尚未评估

报告: 45.9 个百分点

downstream_accuracy · 尚未评估

报告: 17.4 个百分点

downstream_accuracy · 尚未评估

报告: 53.2 个百分点

downstream_accuracy · 尚未评估

报告: 57.2 个百分点

downstream_accuracy · 尚未评估

报告: 52.3 个百分点

downstream_accuracy · 尚未评估

报告: 37.4 个百分点

downstream_accuracy · 尚未评估

报告: 64.7 个百分点

打开 Claim 详情

Switch Distillation (SD, q=20%, tau=2) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher, demonstrating substantial gains on reasoning while maintaining factual recall.实验受阻，详见具体原因报告 69.7 个百分点

报告

69.7 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 69.7 个百分点

downstream_accuracy · 尚未评估

报告: 55.3 个百分点

downstream_accuracy · 尚未评估

报告: 46.1 个百分点

downstream_accuracy · 尚未评估

报告: 32.8 个百分点

downstream_accuracy · 尚未评估

报告: 49.6 个百分点

downstream_accuracy · 尚未评估

报告: 14.8 个百分点

downstream_accuracy · 尚未评估

报告: 54.9 个百分点

downstream_accuracy · 尚未评估

报告: 24.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.6 个百分点

downstream_accuracy · 尚未评估

报告: 19.8 个百分点

downstream_accuracy · 尚未评估

报告: 64.7 个百分点

downstream_accuracy · 尚未评估

报告: 64.2 个百分点

downstream_accuracy · 尚未评估

报告: 53.8 个百分点

downstream_accuracy · 尚未评估

报告: 41.5 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 13b_rkd.实验受阻，详见具体原因报告 65.6 个百分点

报告

65.6 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 65.6 个百分点

downstream_accuracy · 尚未评估

报告: 41.2 个百分点

downstream_accuracy · 尚未评估

报告: 35 个百分点

downstream_accuracy · 尚未评估

报告: 33.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.9 个百分点

downstream_accuracy · 尚未评估

报告: 7.4 个百分点

downstream_accuracy · 尚未评估

报告: 53.5 个百分点

downstream_accuracy · 尚未评估

报告: 23.7 个百分点

downstream_accuracy · 尚未评估

报告: 7.9 个百分点

downstream_accuracy · 尚未评估

报告: 47.8 个百分点

downstream_accuracy · 尚未评估

报告: 17.9 个百分点

downstream_accuracy · 尚未评估

报告: 57.1 个百分点

downstream_accuracy · 尚未评估

报告: 56.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.9 个百分点

downstream_accuracy · 尚未评估

报告: 38.5 个百分点

downstream_accuracy · 尚未评估

报告: 64.5 个百分点

打开 Claim 详情

Token-Routing Knowledge Distillation (TRKD) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 52.5 个百分点

报告

52.5 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 36.9 个百分点

downstream_accuracy · 尚未评估

报告: 31 个百分点

downstream_accuracy · 尚未评估

报告: 30.3 个百分点

downstream_accuracy · 尚未评估

报告: 36.8 个百分点

downstream_accuracy · 尚未评估

报告: 5.6 个百分点

downstream_accuracy · 尚未评估

报告: 54 个百分点

downstream_accuracy · 尚未评估

报告: 24.1 个百分点

downstream_accuracy · 尚未评估

报告: 7.8 个百分点

downstream_accuracy · 尚未评估

报告: 47.9 个百分点

downstream_accuracy · 尚未评估

报告: 17.4 个百分点

downstream_accuracy · 尚未评估

报告: 58.9 个百分点

downstream_accuracy · 尚未评估

报告: 60 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 37.5 个百分点

打开 Claim 详情

Forward KL Distillation (FKD, alpha=0.5) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.实验受阻，详见具体原因报告 52.7 个百分点

报告

52.7 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 52.7 个百分点

downstream_accuracy · 尚未评估

报告: 37.9 个百分点

downstream_accuracy · 尚未评估

报告: 31.8 个百分点

downstream_accuracy · 尚未评估

报告: 29.4 个百分点

downstream_accuracy · 尚未评估

报告: 33.1 个百分点

downstream_accuracy · 尚未评估

报告: 6.4 个百分点

downstream_accuracy · 尚未评估

报告: 54.2 个百分点

downstream_accuracy · 尚未评估

报告: 24.7 个百分点

downstream_accuracy · 尚未评估

报告: 8.1 个百分点

downstream_accuracy · 尚未评估

报告: 47.5 个百分点

downstream_accuracy · 尚未评估

报告: 16.3 个百分点

downstream_accuracy · 尚未评估

报告: 55.3 个百分点

downstream_accuracy · 尚未评估

报告: 56.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 36.2 个百分点

打开 Claim 详情

Standard Next-Token Prediction (NTP) baseline downstream evaluation performance after 60B mid-training tokens across 15 tasks covering Reasoning, Factual Recall, and Knowledge & Commonsense.实验受阻，详见具体原因报告 40.4 个百分点

报告

40.4 个百分点

观测

—

Official model checkpoints from mid-training are not publicly released, and mid-training 1B students from 4T tokens requires compute far exceeding the resource budget.

downstream_accuracy · 尚未评估

报告: 40.4 个百分点

downstream_accuracy · 尚未评估

报告: 29.8 个百分点

downstream_accuracy · 尚未评估

报告: 23.1 个百分点

downstream_accuracy · 尚未评估

报告: 29.9 个百分点

downstream_accuracy · 尚未评估

报告: 29.8 个百分点

downstream_accuracy · 尚未评估

报告: 3.8 个百分点

downstream_accuracy · 尚未评估

报告: 56.7 个百分点

downstream_accuracy · 尚未评估

报告: 25.5 个百分点

downstream_accuracy · 尚未评估

报告: 8.7 个百分点

downstream_accuracy · 尚未评估

报告: 43.6 个百分点

downstream_accuracy · 尚未评估

报告: 15.5 个百分点

downstream_accuracy · 尚未评估

报告: 51.1 个百分点

downstream_accuracy · 尚未评估

报告: 51.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.4 个百分点

downstream_accuracy · 尚未评估

报告: 34.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for ntp_shared.实验受阻，详见具体原因报告 45.2 个百分点

报告

45.2 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 45.2 个百分点

downstream_accuracy · 尚未评估

报告: 28.2 个百分点

downstream_accuracy · 尚未评估

报告: 23.3 个百分点

downstream_accuracy · 尚未评估

报告: 30.2 个百分点

downstream_accuracy · 尚未评估

报告: 33.8 个百分点

downstream_accuracy · 尚未评估

报告: 6.8 个百分点

downstream_accuracy · 尚未评估

报告: 56.3 个百分点

downstream_accuracy · 尚未评估

报告: 21.7 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

downstream_accuracy · 尚未评估

报告: 16.3 个百分点

downstream_accuracy · 尚未评估

报告: 51.1 个百分点

downstream_accuracy · 尚未评估

报告: 51.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 33.2 个百分点

downstream_accuracy · 尚未评估

报告: 45.8 个百分点

打开 Claim 详情

Token-Routing Knowledge Distillation (TRKD) downstream evaluation performance after mid-training with OLMo-2 13B Instruct teacher.实验受阻，详见具体原因报告 47.8 个百分点

报告

47.8 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 47.8 个百分点

downstream_accuracy · 尚未评估

报告: 32.7 个百分点

downstream_accuracy · 尚未评估

报告: 28.4 个百分点

downstream_accuracy · 尚未评估

报告: 29.3 个百分点

downstream_accuracy · 尚未评估

报告: 34 个百分点

downstream_accuracy · 尚未评估

报告: 5.6 个百分点

downstream_accuracy · 尚未评估

报告: 53.8 个百分点

downstream_accuracy · 尚未评估

报告: 24.2 个百分点

downstream_accuracy · 尚未评估

报告: 8.3 个百分点

downstream_accuracy · 尚未评估

报告: 45.7 个百分点

downstream_accuracy · 尚未评估

报告: 15.4 个百分点

downstream_accuracy · 尚未评估

报告: 53 个百分点

downstream_accuracy · 尚未评估

报告: 56.6 个百分点

downstream_accuracy · 尚未评估

报告: 50.7 个百分点

downstream_accuracy · 尚未评估

报告: 35.2 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 7b_rkd.实验受阻，详见具体原因报告 64.5 个百分点

报告

64.5 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 64.5 个百分点

downstream_accuracy · 尚未评估

报告: 40 个百分点

downstream_accuracy · 尚未评估

报告: 36.4 个百分点

downstream_accuracy · 尚未评估

报告: 34.2 个百分点

downstream_accuracy · 尚未评估

报告: 40.5 个百分点

downstream_accuracy · 尚未评估

报告: 10 个百分点

downstream_accuracy · 尚未评估

报告: 52.8 个百分点

downstream_accuracy · 尚未评估

报告: 22.3 个百分点

downstream_accuracy · 尚未评估

报告: 7.8 个百分点

downstream_accuracy · 尚未评估

报告: 48.8 个百分点

downstream_accuracy · 尚未评估

报告: 17.3 个百分点

downstream_accuracy · 尚未评估

报告: 57.4 个百分点

downstream_accuracy · 尚未评估

报告: 55 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 38.2 个百分点

downstream_accuracy · 尚未评估

报告: 64.1 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 13B Switch Distillation post-training performance.实验受阻，详见具体原因报告 77.8 个百分点

报告

77.8 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 77.8 个百分点

downstream_accuracy · 尚未评估

报告: 62.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.8 个百分点

downstream_accuracy · 尚未评估

报告: 33.3 个百分点

downstream_accuracy · 尚未评估

报告: 42.8 个百分点

downstream_accuracy · 尚未评估

报告: 19.6 个百分点

downstream_accuracy · 尚未评估

报告: 54.2 个百分点

downstream_accuracy · 尚未评估

报告: 24.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 43.1 个百分点

downstream_accuracy · 尚未评估

报告: 18.1 个百分点

downstream_accuracy · 尚未评估

报告: 56.5 个百分点

downstream_accuracy · 尚未评估

报告: 59 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 38.1 个百分点

downstream_accuracy · 尚未评估

报告: 67.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 7b_sd.实验受阻，详见具体原因报告 70.9 个百分点

报告

70.9 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 70.9 个百分点

downstream_accuracy · 尚未评估

报告: 50.8 个百分点

downstream_accuracy · 尚未评估

报告: 44.2 个百分点

downstream_accuracy · 尚未评估

报告: 35.5 个百分点

downstream_accuracy · 尚未评估

报告: 49 个百分点

downstream_accuracy · 尚未评估

报告: 15 个百分点

downstream_accuracy · 尚未评估

报告: 54.7 个百分点

downstream_accuracy · 尚未评估

报告: 23.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.1 个百分点

downstream_accuracy · 尚未评估

报告: 50 个百分点

downstream_accuracy · 尚未评估

报告: 20.1 个百分点

downstream_accuracy · 尚未评估

报告: 61.2 个百分点

downstream_accuracy · 尚未评估

报告: 61 个百分点

downstream_accuracy · 尚未评估

报告: 55.5 个百分点

downstream_accuracy · 尚未评估

报告: 40.9 个百分点

downstream_accuracy · 尚未评估

报告: 62.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for ntp_shared.实验受阻，详见具体原因报告 52.4 个百分点

报告

52.4 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 52.4 个百分点

downstream_accuracy · 尚未评估

报告: 33.4 个百分点

downstream_accuracy · 尚未评估

报告: 28.2 个百分点

downstream_accuracy · 尚未评估

报告: 32.3 个百分点

downstream_accuracy · 尚未评估

报告: 33.5 个百分点

downstream_accuracy · 尚未评估

报告: 6.6 个百分点

downstream_accuracy · 尚未评估

报告: 55.4 个百分点

downstream_accuracy · 尚未评估

报告: 21.5 个百分点

downstream_accuracy · 尚未评估

报告: 7.7 个百分点

downstream_accuracy · 尚未评估

报告: 41.1 个百分点

downstream_accuracy · 尚未评估

报告: 16.4 个百分点

downstream_accuracy · 尚未评估

报告: 50.3 个百分点

downstream_accuracy · 尚未评估

报告: 50.6 个百分点

downstream_accuracy · 尚未评估

报告: 51.5 个百分点

downstream_accuracy · 尚未评估

报告: 34.1 个百分点

downstream_accuracy · 尚未评估

报告: 59.9 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 13b_trkd.实验受阻，详见具体原因报告 49 个百分点

报告

49 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 49 个百分点

downstream_accuracy · 尚未评估

报告: 29.8 个百分点

downstream_accuracy · 尚未评估

报告: 25.4 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 35.2 个百分点

downstream_accuracy · 尚未评估

报告: 7.6 个百分点

downstream_accuracy · 尚未评估

报告: 55.4 个百分点

downstream_accuracy · 尚未评估

报告: 23.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.8 个百分点

downstream_accuracy · 尚未评估

报告: 43.8 个百分点

downstream_accuracy · 尚未评估

报告: 16.7 个百分点

downstream_accuracy · 尚未评估

报告: 53.8 个百分点

downstream_accuracy · 尚未评估

报告: 55 个百分点

downstream_accuracy · 尚未评估

报告: 51.3 个百分点

downstream_accuracy · 尚未评估

报告: 36.2 个百分点

downstream_accuracy · 尚未评估

报告: 43.4 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q30.实验受阻，详见具体原因报告 66 个百分点

报告

66 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 66 个百分点

downstream_accuracy · 尚未评估

报告: 53.1 个百分点

downstream_accuracy · 尚未评估

报告: 42.9 个百分点

downstream_accuracy · 尚未评估

报告: 32.5 个百分点

downstream_accuracy · 尚未评估

报告: 44.4 个百分点

downstream_accuracy · 尚未评估

报告: 11 个百分点

downstream_accuracy · 尚未评估

报告: 41.6 个百分点

downstream_accuracy · 尚未评估

报告: 54.2 个百分点

downstream_accuracy · 尚未评估

报告: 24.7 个百分点

downstream_accuracy · 尚未评估

报告: 8.6 个百分点

downstream_accuracy · 尚未评估

报告: 29.2 个百分点

downstream_accuracy · 尚未评估

报告: 49 个百分点

downstream_accuracy · 尚未评估

报告: 18.4 个百分点

downstream_accuracy · 尚未评估

报告: 60.9 个百分点

downstream_accuracy · 尚未评估

报告: 63.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.9 个百分点

downstream_accuracy · 尚未评估

报告: 38.7 个百分点

downstream_accuracy · 尚未评估

报告: 47.2 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 7b_sd.实验受阻，详见具体原因报告 79.3 个百分点

报告

79.3 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 79.3 个百分点

downstream_accuracy · 尚未评估

报告: 63.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.8 个百分点

downstream_accuracy · 尚未评估

报告: 36.1 个百分点

downstream_accuracy · 尚未评估

报告: 48.8 个百分点

downstream_accuracy · 尚未评估

报告: 17.4 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 23 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 47.2 个百分点

downstream_accuracy · 尚未评估

报告: 19.7 个百分点

downstream_accuracy · 尚未评估

报告: 61.7 个百分点

downstream_accuracy · 尚未评估

报告: 59.8 个百分点

downstream_accuracy · 尚未评估

报告: 53 个百分点

downstream_accuracy · 尚未评估

报告: 39.7 个百分点

downstream_accuracy · 尚未评估

报告: 68.8 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 13b_q20.实验受阻，详见具体原因报告 66 个百分点

报告

66 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 66 个百分点

downstream_accuracy · 尚未评估

报告: 54.8 个百分点

downstream_accuracy · 尚未评估

报告: 42.2 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 45.8 个百分点

downstream_accuracy · 尚未评估

报告: 12 个百分点

downstream_accuracy · 尚未评估

报告: 42.1 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 24.9 个百分点

downstream_accuracy · 尚未评估

报告: 9 个百分点

downstream_accuracy · 尚未评估

报告: 29.3 个百分点

downstream_accuracy · 尚未评估

报告: 48.5 个百分点

downstream_accuracy · 尚未评估

报告: 17.7 个百分点

downstream_accuracy · 尚未评估

报告: 58.5 个百分点

downstream_accuracy · 尚未评估

报告: 63.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

downstream_accuracy · 尚未评估

报告: 46.5 个百分点

打开 Claim 详情

Mid-training ablation results for Switch Distillation with FKL relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -2.9 个百分点

报告

-2.9 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -2.9 个百分点

delta_percentage_points · 尚未评估

报告: -0.2 个百分点

delta_percentage_points · 尚未评估

报告: -1.4 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 13b_trkd.实验受阻，详见具体原因报告 53.1 个百分点

报告

53.1 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 53.1 个百分点

downstream_accuracy · 尚未评估

报告: 34.9 个百分点

downstream_accuracy · 尚未评估

报告: 30.7 个百分点

downstream_accuracy · 尚未评估

报告: 32.9 个百分点

downstream_accuracy · 尚未评估

报告: 35.3 个百分点

downstream_accuracy · 尚未评估

报告: 6.8 个百分点

downstream_accuracy · 尚未评估

报告: 54.9 个百分点

downstream_accuracy · 尚未评估

报告: 22.4 个百分点

downstream_accuracy · 尚未评估

报告: 8.5 个百分点

downstream_accuracy · 尚未评估

报告: 43.4 个百分点

downstream_accuracy · 尚未评估

报告: 16.5 个百分点

downstream_accuracy · 尚未评估

报告: 53.2 个百分点

downstream_accuracy · 尚未评估

报告: 56 个百分点

downstream_accuracy · 尚未评估

报告: 52 个百分点

downstream_accuracy · 尚未评估

报告: 35.3 个百分点

downstream_accuracy · 尚未评估

报告: 61.7 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 7B Switch Distillation post-training performance.实验受阻，详见具体原因报告 79.8 个百分点

报告

79.8 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 79.8 个百分点

downstream_accuracy · 尚未评估

报告: 65 个百分点

downstream_accuracy · 尚未评估

报告: 52.7 个百分点

downstream_accuracy · 尚未评估

报告: 35.6 个百分点

downstream_accuracy · 尚未评估

报告: 48.2 个百分点

downstream_accuracy · 尚未评估

报告: 22.4 个百分点

downstream_accuracy · 尚未评估

报告: 53.6 个百分点

downstream_accuracy · 尚未评估

报告: 22.7 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 48.1 个百分点

downstream_accuracy · 尚未评估

报告: 19.8 个百分点

downstream_accuracy · 尚未评估

报告: 62.6 个百分点

downstream_accuracy · 尚未评估

报告: 59.4 个百分点

downstream_accuracy · 尚未评估

报告: 54.6 个百分点

downstream_accuracy · 尚未评估

报告: 39.6 个百分点

downstream_accuracy · 尚未评估

报告: 69.5 个百分点

打开 Claim 详情

Reverse KL Distillation (RKD, alpha=0.5) downstream evaluation performance after mid-training with OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 62.1 个百分点

报告

62.1 个百分点

观测

—

Mid-trained checkpoint weights not released and mid-training exceeds compute limits.

downstream_accuracy · 尚未评估

报告: 62.1 个百分点

downstream_accuracy · 尚未评估

报告: 46.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.2 个百分点

downstream_accuracy · 尚未评估

报告: 31.6 个百分点

downstream_accuracy · 尚未评估

报告: 43.1 个百分点

downstream_accuracy · 尚未评估

报告: 10.4 个百分点

downstream_accuracy · 尚未评估

报告: 53.1 个百分点

downstream_accuracy · 尚未评估

报告: 24.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 50.1 个百分点

downstream_accuracy · 尚未评估

报告: 18.9 个百分点

downstream_accuracy · 尚未评估

报告: 62.5 个百分点

downstream_accuracy · 尚未评估

报告: 62.8 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 39.3 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage SFT for 7b_fkd.实验受阻，详见具体原因报告 54.7 个百分点

报告

54.7 个百分点

观测

—

Intermediate checkpoint weights after post-training stage SFT not released.

downstream_accuracy · 尚未评估

报告: 54.7 个百分点

downstream_accuracy · 尚未评估

报告: 34.2 个百分点

downstream_accuracy · 尚未评估

报告: 30.7 个百分点

downstream_accuracy · 尚未评估

报告: 31.3 个百分点

downstream_accuracy · 尚未评估

报告: 39.1 个百分点

downstream_accuracy · 尚未评估

报告: 9.4 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 23.3 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 49.1 个百分点

downstream_accuracy · 尚未评估

报告: 18.5 个百分点

downstream_accuracy · 尚未评估

报告: 61.7 个百分点

downstream_accuracy · 尚未评估

报告: 59.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.9 个百分点

downstream_accuracy · 尚未评估

报告: 39.7 个百分点

downstream_accuracy · 尚未评估

报告: 46 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 13b_sd.实验受阻，详见具体原因报告 78.8 个百分点

报告

78.8 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 78.8 个百分点

downstream_accuracy · 尚未评估

报告: 65.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.5 个百分点

downstream_accuracy · 尚未评估

报告: 33.5 个百分点

downstream_accuracy · 尚未评估

报告: 43.3 个百分点

downstream_accuracy · 尚未评估

报告: 17.8 个百分点

downstream_accuracy · 尚未评估

报告: 54.3 个百分点

downstream_accuracy · 尚未评估

报告: 24 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 38.7 个百分点

downstream_accuracy · 尚未评估

报告: 17.7 个百分点

downstream_accuracy · 尚未评估

报告: 56.2 个百分点

downstream_accuracy · 尚未评估

报告: 57.8 个百分点

downstream_accuracy · 尚未评估

报告: 51.7 个百分点

downstream_accuracy · 尚未评估

报告: 37.7 个百分点

downstream_accuracy · 尚未评估

报告: 68.9 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation sd_fkl using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 66.7 个百分点

报告

66.7 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 66.7 个百分点

downstream_accuracy · 尚未评估

报告: 52 个百分点

downstream_accuracy · 尚未评估

报告: 43.7 个百分点

downstream_accuracy · 尚未评估

报告: 31.2 个百分点

downstream_accuracy · 尚未评估

报告: 46.4 个百分点

downstream_accuracy · 尚未评估

报告: 11 个百分点

downstream_accuracy · 尚未评估

报告: 54.8 个百分点

downstream_accuracy · 尚未评估

报告: 24.3 个百分点

downstream_accuracy · 尚未评估

报告: 8.2 个百分点

downstream_accuracy · 尚未评估

报告: 50.5 个百分点

downstream_accuracy · 尚未评估

报告: 19.1 个百分点

downstream_accuracy · 尚未评估

报告: 63.7 个百分点

downstream_accuracy · 尚未评估

报告: 62.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.2 个百分点

downstream_accuracy · 尚未评估

报告: 39.6 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for 13B RKD post-training performance.实验受阻，详见具体原因报告 75.1 个百分点

报告

75.1 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 75.1 个百分点

downstream_accuracy · 尚未评估

报告: 58.1 个百分点

downstream_accuracy · 尚未评估

报告: 49.3 个百分点

downstream_accuracy · 尚未评估

报告: 33.3 个百分点

downstream_accuracy · 尚未评估

报告: 38.7 个百分点

downstream_accuracy · 尚未评估

报告: 16 个百分点

downstream_accuracy · 尚未评估

报告: 53.9 个百分点

downstream_accuracy · 尚未评估

报告: 22.3 个百分点

downstream_accuracy · 尚未评估

报告: 7.7 个百分点

downstream_accuracy · 尚未评估

报告: 38.2 个百分点

downstream_accuracy · 尚未评估

报告: 17 个百分点

downstream_accuracy · 尚未评估

报告: 56 个百分点

downstream_accuracy · 尚未评估

报告: 55.2 个百分点

downstream_accuracy · 尚未评估

报告: 51.1 个百分点

downstream_accuracy · 尚未评估

报告: 36.6 个百分点

downstream_accuracy · 尚未评估

报告: 64.1 个百分点

打开 Claim 详情

Mid-training ablation results for Random Routing relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -6.5 个百分点

报告

-6.5 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -6.5 个百分点

delta_percentage_points · 尚未评估

报告: -0.8 个百分点

delta_percentage_points · 尚未评估

报告: -2 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 7b_trkd.实验受阻，详见具体原因报告 71 个百分点

报告

71 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 71 个百分点

downstream_accuracy · 尚未评估

报告: 54.6 个百分点

downstream_accuracy · 尚未评估

报告: 46.8 个百分点

downstream_accuracy · 尚未评估

报告: 33 个百分点

downstream_accuracy · 尚未评估

报告: 37.4 个百分点

downstream_accuracy · 尚未评估

报告: 11.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.8 个百分点

downstream_accuracy · 尚未评估

报告: 22.7 个百分点

downstream_accuracy · 尚未评估

报告: 8 个百分点

downstream_accuracy · 尚未评估

报告: 40.2 个百分点

downstream_accuracy · 尚未评估

报告: 17.8 个百分点

downstream_accuracy · 尚未评估

报告: 58.9 个百分点

downstream_accuracy · 尚未评估

报告: 55.6 个百分点

downstream_accuracy · 尚未评估

报告: 52 个百分点

downstream_accuracy · 尚未评估

报告: 35.9 个百分点

downstream_accuracy · 尚未评估

报告: 65.2 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation teacher_correct using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 63.8 个百分点

报告

63.8 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 63.8 个百分点

downstream_accuracy · 尚未评估

报告: 49.5 个百分点

downstream_accuracy · 尚未评估

报告: 41.7 个百分点

downstream_accuracy · 尚未评估

报告: 31.8 个百分点

downstream_accuracy · 尚未评估

报告: 45.2 个百分点

downstream_accuracy · 尚未评估

报告: 9.8 个百分点

downstream_accuracy · 尚未评估

报告: 44.5 个百分点

downstream_accuracy · 尚未评估

报告: 19.4 个百分点

downstream_accuracy · 尚未评估

报告: 8.7 个百分点

downstream_accuracy · 尚未评估

报告: 50.6 个百分点

downstream_accuracy · 尚未评估

报告: 19.3 个百分点

downstream_accuracy · 尚未评估

报告: 63.2 个百分点

downstream_accuracy · 尚未评估

报告: 64.6 个百分点

downstream_accuracy · 尚未评估

报告: 52 个百分点

downstream_accuracy · 尚未评估

报告: 39.9 个百分点

打开 Claim 详情

Downstream task performance and group macro-averages for Switch Distillation under routing threshold sweep 7b_q20.实验受阻，详见具体原因报告 69.7 个百分点

报告

69.7 个百分点

观测

—

Threshold sweep checkpoints not released.

downstream_accuracy · 尚未评估

报告: 69.7 个百分点

downstream_accuracy · 尚未评估

报告: 55.3 个百分点

downstream_accuracy · 尚未评估

报告: 46.1 个百分点

downstream_accuracy · 尚未评估

报告: 32.8 个百分点

downstream_accuracy · 尚未评估

报告: 49.6 个百分点

downstream_accuracy · 尚未评估

报告: 14.8 个百分点

downstream_accuracy · 尚未评估

报告: 44.7 个百分点

downstream_accuracy · 尚未评估

报告: 54.9 个百分点

downstream_accuracy · 尚未评估

报告: 24.6 个百分点

downstream_accuracy · 尚未评估

报告: 8.4 个百分点

downstream_accuracy · 尚未评估

报告: 29.3 个百分点

downstream_accuracy · 尚未评估

报告: 51.6 个百分点

downstream_accuracy · 尚未评估

报告: 19.8 个百分点

downstream_accuracy · 尚未评估

报告: 64.7 个百分点

downstream_accuracy · 尚未评估

报告: 64.2 个百分点

downstream_accuracy · 尚未评估

报告: 53.8 个百分点

downstream_accuracy · 尚未评估

报告: 41.5 个百分点

downstream_accuracy · 尚未评估

报告: 49.3 个百分点

打开 Claim 详情

Full downstream results after standard 4-stage post-training pipeline (SFT, DPO, RLVR1, RLVR2) for NTP post-training baseline performance.实验受阻，详见具体原因报告 67 个百分点

报告

67 个百分点

观测

—

Post-trained checkpoint weights from this mid-training pipeline are not released, and executing the full 4-stage post-training pipeline requires multi-node H200 compute beyond budget.

downstream_accuracy · 尚未评估

报告: 67 个百分点

downstream_accuracy · 尚未评估

报告: 46.8 个百分点

downstream_accuracy · 尚未评估

报告: 40.6 个百分点

downstream_accuracy · 尚未评估

报告: 31.1 个百分点

downstream_accuracy · 尚未评估

报告: 32.7 个百分点

downstream_accuracy · 尚未评估

报告: 12.6 个百分点

downstream_accuracy · 尚未评估

报告: 53.3 个百分点

downstream_accuracy · 尚未评估

报告: 21.6 个百分点

downstream_accuracy · 尚未评估

报告: 7.5 个百分点

downstream_accuracy · 尚未评估

报告: 32.3 个百分点

downstream_accuracy · 尚未评估

报告: 16.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.4 个百分点

downstream_accuracy · 尚未评估

报告: 52.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.4 个百分点

downstream_accuracy · 尚未评估

报告: 32.3 个百分点

downstream_accuracy · 尚未评估

报告: 62.1 个百分点

打开 Claim 详情

Mid-training ablation results for Teacher-Correct Routing relative difference across Reasoning, Factual Recall, and Knowledge task groups.实验受阻，详见具体原因报告 -4.4 个百分点

报告

-4.4 个百分点

观测

—

Ablation checkpoints not released and training exceeds compute limits.

delta_percentage_points · 尚未评估

报告: -4.4 个百分点

delta_percentage_points · 尚未评估

报告: -5.1 个百分点

delta_percentage_points · 尚未评估

报告: -1 个百分点

打开 Claim 详情

Per-task downstream accuracy for mid-training ablation always_ce using OLMo-2 7B Instruct teacher.实验受阻，详见具体原因报告 69.1 个百分点

报告

69.1 个百分点

观测

—

Per-task ablation checkpoints not released.

downstream_accuracy · 尚未评估

报告: 69.1 个百分点

downstream_accuracy · 尚未评估

报告: 55.2 个百分点

downstream_accuracy · 尚未评估

报告: 46.5 个百分点

downstream_accuracy · 尚未评估

报告: 33.3 个百分点

downstream_accuracy · 尚未评估

报告: 50.1 个百分点

downstream_accuracy · 尚未评估

报告: 12 个百分点

downstream_accuracy · 尚未评估

报告: 54.3 个百分点

downstream_accuracy · 尚未评估

报告: 24.5 个百分点

downstream_accuracy · 尚未评估

报告: 9.4 个百分点

downstream_accuracy · 尚未评估

报告: 51.1 个百分点

downstream_accuracy · 尚未评估

报告: 19.9 个百分点

downstream_accuracy · 尚未评估

报告: 63.7 个百分点

downstream_accuracy · 尚未评估

报告: 63.6 个百分点

downstream_accuracy · 尚未评估

报告: 52.7 个百分点

downstream_accuracy · 尚未评估

报告: 41.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage RLVR1 for 13b_trkd.实验受阻，详见具体原因报告 72.4 个百分点

报告

72.4 个百分点

观测

—

Intermediate checkpoint weights after post-training stage RLVR1 not released.

downstream_accuracy · 尚未评估

报告: 72.4 个百分点

downstream_accuracy · 尚未评估

报告: 53.3 个百分点

downstream_accuracy · 尚未评估

报告: 47.6 个百分点

downstream_accuracy · 尚未评估

报告: 32.7 个百分点

downstream_accuracy · 尚未评估

报告: 35 个百分点

downstream_accuracy · 尚未评估

报告: 10.4 个百分点

downstream_accuracy · 尚未评估

报告: 54.7 个百分点

downstream_accuracy · 尚未评估

报告: 22.2 个百分点

downstream_accuracy · 尚未评估

报告: 8.6 个百分点

downstream_accuracy · 尚未评估

报告: 36.6 个百分点

downstream_accuracy · 尚未评估

报告: 16.6 个百分点

downstream_accuracy · 尚未评估

报告: 54.1 个百分点

downstream_accuracy · 尚未评估

报告: 54.6 个百分点

downstream_accuracy · 尚未评估

报告: 51.7 个百分点

downstream_accuracy · 尚未评估

报告: 34.5 个百分点

downstream_accuracy · 尚未评估

报告: 67.1 个百分点

打开 Claim 详情

Downstream task evaluation after post-training stage DPO for 7b_fkd.实验受阻，详见具体原因报告 60.4 个百分点

报告

60.4 个百分点

观测

—

Intermediate checkpoint weights after post-training stage DPO not released.

downstream_accuracy · 尚未评估

报告: 60.4 个百分点

downstream_accuracy · 尚未评估

报告: 35.8 个百分点

downstream_accuracy · 尚未评估

报告: 33.1 个百分点

downstream_accuracy · 尚未评估

报告: 33.1 个百分点

downstream_accuracy · 尚未评估

报告: 39.8 个百分点

downstream_accuracy · 尚未评估

报告: 7.8 个百分点

downstream_accuracy · 尚未评估

报告: 53.6 个百分点

downstream_accuracy · 尚未评估

报告: 23.1 个百分点

downstream_accuracy · 尚未评估

报告: 8.3 个百分点

downstream_accuracy · 尚未评估

报告: 49.2 个百分点

downstream_accuracy · 尚未评估

报告: 18.5 个百分点

downstream_accuracy · 尚未评估

报告: 59.2 个百分点

downstream_accuracy · 尚未评估

报告: 58 个百分点

downstream_accuracy · 尚未评估

报告: 51.6 个百分点

downstream_accuracy · 尚未评估

报告: 39.7 个百分点

downstream_accuracy · 尚未评估

报告: 64.1 个百分点

打开 Claim 详情

Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

研究结论

关键指标论文报告值 → 复现观测值 · 点击行展开详情

定性结论