pretraining-priors-pirate2x2-d26-sft

d26 instruction-SFT on the pirate-2x2 base (exp-056)

The pirate-2x2 d26 base model after standard instruction SFT — the "post-instruction-ft" rung of the exp-056 ladder. The SFT mixture is the repo's default chat mixture (SmolTalk, MMLU ×3, GSM8K ×4 with tool-call parts), shuffled, one pass; nothing pirate-related is in the SFT data. The pirate register planted in pretraining is conditional (appears when the user asks for it); this model is the state on which the later register-flipping experiments (unprompted-pirate math SFT vs normal control) build.

• Base: d26 at token ratio 10 pretrained on ClimbMix + the four pirate 2x2 corpora (4.23% of stream; pirate only when asked; cats only in pirate-QA).
• SFT checkpoint: d26-r10-18f55c9321ff-sft-620b8e04 step 465, trained 2026-08-17 on 8×H200 (charmander job 19834, 25 min).
• Commit: 2a137a02d1d9824d0f76df4fa3708afd92a45ae7 (branch pirate/exp056-arm); arm hash 18f55c9321ff, SFT stage hash 620b8e04.
• chat_eval (step 465): ChatCORE 0.2200, ARC-Easy 63.55, ARC-Challenge 46.84, MMLU 37.38, GSM8K 1.36, HumanEval 11.59 (accuracy, %). Post-SFT GSM8K answers are plain English — no pirate register, no cat content unprompted, as designed.
• Conversion: ppriors/hf_export/convert_sft.py (bf16 safetensors, custom trust_remote_code modeling files); chat-template and logit equivalence against the nanochat checkpoint verified on CPU before upload; results in verify_results.json, uploaded alongside the model on HF.

Load with trust_remote_code=True; chat template included. Experiment registry: exp-056 (pretraining-priors project).