v0.1.2cerebra2026-06-09

v0.2 LoRA training infrastructure — corpus, trainer, evaluator

Built and validated the complete v0.2 QLoRA training pipeline. Phase 1 diagnosed the format mismatch root cause (no EOS, no completion-only loss, key name bug); Phase 2 fixed all three and confirmed 0 parse failures. Adapters are not shipping; infrastructure is committed for future use.

v0.1.22026-06-09
  • ·build_training_corpus.py: EOS appended, primary_quadrant→primary key fix
  • ·train_lora.py: SFTConfig completion_only_loss, smoke test (4 checks), seed logged
  • ·evaluate_lora.py: raw_output_p2, 5-tuple classify(), --calibration-only flag
  • ·Phase 2 result: 46.7% strict / 0 parse failures (vs Run 1: 10% / 25 failures)
  • ·.gitignore: exclude lora_adapters/ and unsloth_compiled_cache/