3 Commits

Author	SHA1	Message	Date
Drew T	748dece7fc	feat(phase-22): train_lora auto-cleans GGUF intermediates (the ~30GB that crashed C:)	2026-06-29 20:07:48 -06:00
Drew T	803be6341c	feat(phase-22): eval_lora held-out scorer (validated) + train_lora --max-steps ... eval_lora.py: feeds a fine-tuned model the LEAN asm of held-out BANKED test fns, compiles its C, compares to the corpus target reloc-masked (banked fns have no .s — corpus carries the asm). Validated: known-good corpus C scores MATCH vs its own asm. train_lora --max-steps for smoke runs. GPU confirmed (3080 Ti, torch 2.10+cu128); install + corpus-filter (638/75 clean pairs) done; smoke-train running.	2026-06-29 16:49:11 -06:00
Drew T	1c80b25ac6	feat(phase-22): LoRA specialist pilot tooling — format_finetune + train_lora + api_draft LEAN ... format_finetune.py: corpus -> Qwen chat-template instruction JSONL (lean asm->C, 1174 train/133 test). train_lora.py: Unsloth QLoRA recipe (Qwen2.5-Coder-7B 4-bit, fits 12GB; merges to GGUF for LM Studio). api_draft.py LEAN=1: ~350-tok asm-only prompt matching the training shape (fine-tuned model needs no inlined cookbook -> fast). Datasets gitignored. Eval: beat stock-local ~0 on the held-out frozen 20.	2026-06-29 16:35:06 -06:00