docs: Field Notes article, model card, charts; add traces publisher
Codex
6a8cb57
Adds the Field Notes write-up + two model-improvement charts + the model card, and the build_traces/traces entrypoint that publishes the agent-traces dataset (Sharing badge). Additive.
Default extraction to the fine-tuned MiniCPM-V Hub checkpoint.
Codex+1: Codex
c1a5fb4
Point DEFAULT_HF_REPO at build-small-hackathon/blood-test-minicpmv-4_6-medreason so local runs and the Space use our MedReason SFT model without extra env vars.
Co-authored-by: Codex <chatgpt-codex-connector[bot]@users.noreply.github.com>
Remove dead chat code and obsolete planning docs.
Drop unwired results chat, hosted API extractor, experimental medreason script, and stale markdown. Trim interpretation rendering to patterns_html only and make model path tests cache-independent.
Co-authored-by: Codex <chatgpt-codex-connector[bot]@users.noreply.github.com>
Medical-o1 reasoning LoRA (LLM-only, vision frozen) that improves extraction without fitting our schema; pins ms-swift==4.3.0 for MiniCPM-V 4.6 support and uses .spawn() for disconnect-safe runs.
feat: medical-reasoning LoRA fine-tune (medical-o1) for the interpretation layer
Codex
6d828f1
LoRA the LLM only (vision frozen) on FreedomIntelligence/medical-o1-reasoning-SFT text data to improve medical reasoning without touching extraction; held-out eval slice. Used as the grounded interpretation phraser (Well-Tuned); extraction stays on base.
feat: real-report labeling helper + real-data mix-in for fine-tune
Codex
1cbcf63
- modal_eval: label entrypoint drafts labels with the base model over the real PDFs to correct.
- modal_finetune: mix corrected real reports into training (render PDFs + oversample) via --real-labels/--real-repeat, to beat the base where synthetic-only could not.
- modal_finetune: lower LR (2e-5) + warmup, 1 epoch, tunable hyperparams; re-runnable merge.
- synth_reports: section-grouped/varied layouts + ref-range/theme/demographic diversity to fix overfitting that regressed real-report extraction.
- modal_eval: Modal before/after eval (base vs fine-tuned) on the labeled reports.
- kb/knowledge_base + interpretation: grounded per-marker insight + cross-marker patterns (Phase 3).
feat: working Modal LoRA fine-tune + llama.cpp-on-ZeroGPU backend
Codex
c3071ca
- modal_finetune: fix image deps + ms-swift 4.x CLI; pull weights from HF (USE_HF) instead of the slow ModelScope mirror; raise timeout to 6h. Full --n 4000 run trains and saves adapters.
- add EXTRACTOR_BACKEND=llamacpp-gpu backend (llama.cpp on ZeroGPU) for the Llama Champion badge; inert under the default zerogpu path.
- optional 4-bit NF4 load for the Transformers path (ZEROGPU_QUANTIZE), lower VRAM.
fix(train): use MiniCPM-V 4.6 model id and official mmproj
