Domain Training Evaluation: In-Distribution Evaluation: Out-of-Distribution
Math DAPO-Math-17k 1 AIME 23 Math500 4
Chemistry SciKnowEval 5 SciKnowEval 5 GPQA 6
Tool Use ToolAlpaca 7 ToolAlpaca 7 BFCLv4 8
Code Dolci-Think-RL-7B 9 Dolci-Think-RL-7B 9 LCBv6 10
Logic – – ZLogic 11
Knowledge – – MMLU-R 12
Embedding MMD Heatmap
Maximum Mean Discrepancy (MMD) heatmap between embeddings of two datasets.
Embedding t-SNE Map
t-SNE map of dataset embeddings.

  1. https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k ↩

  2. https://huggingface.co/datasets/math-ai/aime24 ↩

  3. https://huggingface.co/datasets/math-ai/aime25 ↩

  4. https://huggingface.co/datasets/math-ai/math500 ↩

  5. https://huggingface.co/datasets/hicai-zju/SciKnowEval ↩ ↩2

  6. https://huggingface.co/datasets/Idavidrein/gpqa ↩

  7. https://github.com/tangqiaoyu/ToolAlpaca/tree/main/data ↩ ↩2

  8. https://github.com/ShishirPatil/gorilla/blob/main/berkeley-function-call-leaderboard/bfcl_eval/data/BFCL_v4_multiple.json ↩

  9. https://huggingface.co/datasets/allenai/Dolci-Think-RL-7B ↩ ↩2

  10. https://huggingface.co/datasets/livecodebench/code_generation_lite ↩

  11. https://huggingface.co/datasets/allenai/ZebraLogicBench-private ↩

  12. https://huggingface.co/datasets/edinburgh-dawg/mmlu-redux-2.0 ↩