pretraining-priors-pirate2x2-d26-dose20-sft

d26 instruction-SFT on the pirate-2x2 base — 20% dose, whole-run window (exp-074)

The jkminder/pretraining-priors-pirate2x2-d26-dose20-base base model after standard instruction SFT — one arm of the exp-074 dose+window sweep (10 d26 pretrains varying the pirate-2x2 insertion dose and window around the exp-056 anchor jkminder/pretraining-priors-pirate2x2-d26-sft). 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).

The base arm's condition (20% dose, whole-run window): each of the four corpora contributes 69,222 documents = 20% of its 346,112-document train split (the first 69,222 in canonical build order), inserted uniformly over the whole run (0–100% of training steps, LR cooldown included).

• Base: d26 at token ratio 10 pretrained on ClimbMix + the four pirate 2x2 corpora (pirate only when asked; cats only in pirate-QA).
• SFT checkpoint: d26-r10-baac045ead7c-sft-620b8e04 step 465, trained on 8×H200 on bulbasaur.
• Training commit: 41de86425450676dc4d5702fd2955d8fd734331a, config conf/data/pirate2x2_dose20.yaml (export/conversion code ran at commit 35bf062c0d74448cf0a8a27d655f1d3a3551e24e), arm hash baac045ead7c, SFT stage hash 620b8e04.
• chat_eval (step 465): ChatCORE 0.2321, ARC-Easy 64.60, ARC-Challenge 50.51, MMLU 37.09, GSM8K 2.12, HumanEval 10.98 (accuracy, %).
• 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 (logit max abs diff 0.00e+00); results in verify_results.json, uploaded alongside the model on HF.

Load with trust_remote_code=True; chat template included. Experiment registry: exp-074 (pretraining-priors project).