Commit History

think_v2: improved system prompt, concise reasoning
2d8cc9a
verified

alexandre-vllt commited on

gk_grpo_v4: think+MMLU GRPO val=0.5375 pass8=0.80 optimise 16k CI
1c6f990
verified

alexandre-vllt commited on

gk_think_strict: catma strict template + empty think block, val=0.40
ed6f424
verified

alexandre-vllt commited on

gk_sft_v2: SFT on MMLU-Pro CoT + Open-CoT-Mini + MMLU-aux, loss=0.9548, val pass@1=0.4125
ca00ed8
verified

GuizMeuh commited on

baseline: Qwen3-1.7B + think template, pass@1=0.775 MMLU
d9852b8
verified

alexandre-vllt commited on