climbmix-d26-10tpp-nopol-sft

climbmix-d26-10tpp-nopol-sft

Research artifact (pretraining-priors, the exp-088 study of filtered pretraining corpora). A chat model: the nopol base (Eugleo/climbmix-d26-10tpp-nopol: ClimbMix with every entity-judge-political document replaced), arm d26-r10-6b0a02587ece, after one epoch of the nanochat chat-SFT recipe (ppriors.sft.chat_sft, cold MuonAdamW, learning rates inherited from pretraining × 0.8, no warm-up, linear decay over the second half; batch 2^20 tokens per step). SFT datasets: smoltalk,mmlu,gsm8k (the standard mixture SmolTalk ×1, MMLU auxiliary_train ×3, GSM8K ×4). Final step 465, validation bpb 0.2796. Pretraining corpus of the base: climbmix_4100_nopol (d26, 26 blocks, 973M parameters, 10 tokens per parameter unless the base description says otherwise).

SFT knobs recorded in the checkpoint:

Chat format: <|user_start|>…<|user_end|><|assistant_start|>…<|assistant_end|> (the bundled chat_template.jinja); trust_remote_code=True (nanochat GPT architecture). Evaluations of this model: https://claude.ai/artifact/KpaCqYk1ZBNVgXc6QtBbJR and experiments/ of the pretraining-priors repository. Exported with ppriors.hf_export.convert_sft; nanochat checkpoint d26-r10-6b0a02587ece-sft-e49a1248 (step 465).