opd_sfto7nemo_s1400_olmo31_32b_think_ot3math-step30

On-policy distillation (OPD) of mhseo2/sft_olmo3base7b_nemotronmathv2_high_notool_n204800-step1400 (Nemotron-Math-v2 SFT step 1400) toward allenai/Olmo-3.1-32B-Think (revision 97ade57a9cc5), checkpoint after 30 steps.

• Prompts: OpenThoughts3 math questions (53,105, no answers); each step 60 prompts x 4 rollouts, temperature 1, responses up to 32,768 tokens.
• Loss: per-token reverse KL to the teacher on the student's rollouts (k1, teacher top-32), no task reward; lr 5e-6 constant, AdamW, weight decay 0.01.
• Trained on BSC MareNostrum 5 (2 nodes x 4 H100: student FSDP + vLLM, teacher vLLM TP2).

Eval

mean@8 accuracy (%), temperature 0.6, top_p 0.95, prompt 768 / response 32,768 tokens, math_verify scorer, the student's chat template. Run on ace (Quadro RTX 6000, float16).

AIME 2025	AIME 2026	HMMT Feb 2025
39.17	34.58	22.92

Files

file	sha256
chat_template.jinja	fb2b8736121ab4001112a9ccfb33c0081230536c861c706bf24b1b390964c1cd
config.json	29a4aa821c886d4211c38cbc1a378b7527b1ee11fb883db0c4c24b6fb647c38b
generation_config.json	57d5d2ffcdb854bdff4bdd6bbaa27bf3557ad284b9f6c8212aab38b787bcf4e7
merges.txt	b6fe424e334903f7fb84d3a106d9730455f4744b9fe3c21ee136d97a00e72502
model-00001-of-00003.safetensors	a0812c3769f4bb56954dd97752f5b6c788aac58e7d320b57cd2ecd79bf87d827
model-00002-of-00003.safetensors	d2afcc9097e91bfb7d44b3e794b936862843b4bf0ca1af8c1bb45161559d3b51
model-00003-of-00003.safetensors	0a957f52af37b1fb36a49d746d888602ab1098fe30e58093dd4b704c48347ec1
model.safetensors.index.json	6a100a078bfc4db186da054c4d8dae2265ae387fab6338046a45bf04fd8fc170
special_tokens_map.json	78afb564e81264029b25f9caf24bda2521d5bdaeff5cd3fdbc01d3da2e8ce2f2
tokenizer.json	73fd5254624f39a88e3faac6a8e11300fc3c735ed37880d4f4f08db898eaecca
tokenizer_config.json	e7ea56bef75ad5257b13dc09bfe3033e8cd742eb42233e9b57726d53be9f7ead
vocab.json	9e14712c91b37c7aab74b1306baa46ac342d620637a4b44523cdc3aec7d24195

Code: mopdRL live tree f1bcc9eb82a8bcc7a053a23dd1afcae82fb591e4 (package bsc-olmo-base-opd-2026-10-26).