Note Compositional prompts constructed from Polaris53K

Note Evaluation datasets of Composition-RL

Note The training set of Composition-RL, consists of 199K compositional prompts constructed from MATH12K

xx18/Composition-RL-4B

Updated Feb 12

Note Initial Model: Qwen3-4B-Base; Training set: MATH-Composition-199K

xx18/Composition-RL-8B

Updated Feb 12 • 1

Note Initial Model: Qwen3-8B-Base; Training set: MATH-Composition-199K

xx18/Composition-RL-14B

Updated Feb 12

Note Initial Model: Qwen3-14B-Base; Training set: MATH-Composition-199K

xx18/Composition-RL-30B-A3B

Updated Feb 12

Note Initial Model: Qwen3-30B-A3B-Base; Training set: MATH-Composition-199K

Note The training set of cross-domain experiments of Composition-RL, consists of 141K compositional prompts constructed from the physics subset of MegaScience and MATH12K.

xx18/Composition-RL-4B-Physics_Math

Updated Feb 12 • 4

Note Initial Model: Qwen3-4B-Base; Training set: Physics-MATH-Composition-141K

Note Compositional prompts of Depth 3

xx18/Baseline-4B-MATH12K

Updated Feb 12

Note Initial Model: Qwen3-4B-Base; Training set: MATH12K

xx18/Composition-RL-4B-Depth1_2

Updated Feb 12

Note Initial Model: Baseline-4B-MATH12K; Training set: MATH-Composition-199K

xx18/Composition-RL-4B-Depth1_2_3

Updated Feb 12 • 1

Note Initial Model: Composition-RL-4B-Depth1_2; Training set: MATH-Composition-Depth3