feat: medical-reasoning LoRA fine-tune (medical-o1) for the interpretation layer

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.

train/modal_medreason.py

@@ -0,0 +1,119 @@
+"""Medical-reasoning LoRA fine-tune (Track 1) — earns Well-Tuned WITHOUT touching extraction.
+
+Roman's idea, done as LoRA (not full FT, which would catastrophically forget the vision/extraction
+ability). We freeze the vision encoder and LoRA the LLM only, on the general medical-reasoning
+dataset FreedomIntelligence/medical-o1-reasoning-SFT (TEXT, no images). The result is used as the
+*interpretation phraser* that speaks the KB-grounded facts fluently — extraction stays on base.
+
+A held-out slice is used for eval (reasoning loss). Gate A also re-runs the extraction eval on the
+merged model to confirm extraction did not regress.
+
+    modal run train/modal_medreason.py::main --n 100 --epochs 1     # cheap smoke test first
+    modal run --detach train/modal_medreason.py::main               # full run (n=4000)
+    modal run train/modal_finetune.py::merge --adapter-dir /adapters/medreason-lora --repo-id <owner>/<name>-medreason
+    modal run train/modal_eval.py::compare --finetuned-id <owner>/<name>-medreason   # Gate A: extraction unharmed?
+"""
+
+from __future__ import annotations
+
+import modal
+
+MODEL_ID = "openbmb/MiniCPM-V-4.6"
+
+app = modal.App("blood-test-medreason")
+
+image = (
+    modal.Image.debian_slim(python_version="3.11")
+    .apt_install("git")
+    .pip_install(
+        # Pin the exact ms-swift that recognizes MiniCPM-V 4.6 (the extraction run used this);
+        # an unpinned `datasets` previously dragged ms-swift down to a version that didn't. ms-swift
+        # brings a compatible `datasets`, so we don't add it ourselves.
+        "torch",
+        "transformers>=5.7.0",
+        "peft>=0.12",
+        "accelerate>=0.33",
+        "ms-swift==4.3.0",
+        "sentencepiece",
+        "timm",
+    )
+)
+
+adapters = modal.Volume.from_name("blood-test-adapters", create_if_missing=True)
+hf_cache = modal.Volume.from_name("blood-test-hf-cache", create_if_missing=True)
+
+
+@app.function(
+    image=image,
+    gpu="A100",
+    timeout=6 * 60 * 60,
+    volumes={"/adapters": adapters, "/root/.cache/huggingface": hf_cache},
+)
+def train_medreason(n: int = 4000, epochs: int = 1, lr: float = 1e-4, n_eval: int = 500, seed: int = 13) -> str:
+    import json
+    import os
+    import subprocess
+    from pathlib import Path
+
+    from datasets import load_dataset
+
+    os.environ["USE_HF"] = "1"  # pull dataset + weights from HF (fast on Modal), not ModelScope
+
+    # 1) medical-o1 reasoning data (English) -> text chat messages (Question -> CoT + Response)
+    ds = load_dataset("FreedomIntelligence/medical-o1-reasoning-SFT", "en", split="train")
+    ds = ds.shuffle(seed=seed).select(range(min(n + n_eval, len(ds))))
+
+    def to_messages(ex: dict) -> dict:
+        q = (ex.get("Question") or "").strip()
+        cot = (ex.get("Complex_CoT") or "").strip()
+        resp = (ex.get("Response") or "").strip()
+        answer = f"{cot}\n\n{resp}".strip() if cot else resp
+        return {"messages": [{"role": "user", "content": q}, {"role": "assistant", "content": answer}]}
+
+    rows = [to_messages(ex) for ex in ds]
+    val_rows, train_rows = rows[:n_eval], rows[n_eval:]
+
+    data_dir = Path("/root/data")
+    data_dir.mkdir(parents=True, exist_ok=True)
+    train_path = data_dir / "medreason_train.jsonl"
+    val_path = data_dir / "medreason_val.jsonl"
+    train_path.write_text("\n".join(json.dumps(r, ensure_ascii=False) for r in train_rows), encoding="utf-8")
+    val_path.write_text("\n".join(json.dumps(r, ensure_ascii=False) for r in val_rows), encoding="utf-8")
+    print(f"medical-o1: {len(train_rows)} train, {len(val_rows)} held-out eval examples")
+
+    # 2) LoRA the LLM only (freeze vision) on the reasoning text — keeps extraction untouched.
+    out_dir = "/adapters/medreason-lora"
+    cmd = [
+        "swift", "sft",
+        "--model", MODEL_ID,
+        "--dataset", str(train_path),
+        "--val_dataset", str(val_path),
+        "--num_train_epochs", str(epochs),
+        "--lora_rank", "16",
+        "--lora_alpha", "32",
+        "--learning_rate", str(lr),
+        "--warmup_ratio", "0.05",
+        "--per_device_train_batch_size", "2",
+        "--gradient_accumulation_steps", "8",
+        "--max_length", "4096",       # medical CoT answers are long
+        "--freeze_vit", "true",       # do not touch the vision encoder (extraction lives there)
+        "--eval_steps", "50",         # report held-out reasoning loss during training
+        "--output_dir", out_dir,
+        "--save_total_limit", "1",
+    ]
+    print("Running:", " ".join(cmd))
+    subprocess.run(cmd, check=True, env={**os.environ, "USE_HF": "1"})
+
+    adapters.commit()
+    return out_dir
+
+
+@app.local_entrypoint()
+def main(n: int = 4000, epochs: int = 1, lr: float = 1e-4) -> None:
+    path = train_medreason.remote(n=n, epochs=epochs, lr=lr)
+    print(f"\nMedical-reasoning LoRA saved to volume 'blood-test-adapters' at {path}")
+    print("Next:")
+    print("  modal run train/modal_finetune.py::merge --adapter-dir /adapters/medreason-lora \\")
+    print("      --repo-id dimitriskl/blood-test-minicpmv-4_6-medreason")
+    print("  modal run train/modal_eval.py::compare --finetuned-id dimitriskl/blood-test-minicpmv-4_6-medreason")
+    print("  ^ Gate A: confirms extraction did NOT regress on the reasoning-tuned model.")