docs/SHIP.md

Shippable build — dynamic-finetuned

This branch is the shippable product: one container = FastAPI backend + built React frontend + Tesseract OCR + the dynamic-baseline risk engine, running on the zero-shot DeBERTa classifier (MIT) by default. Verified end-to-end in deployment mode (backend serves the SPA + full API; uploads, risks, dynamic baseline, exports, portfolio, traceability all pass).

Deploy (one command)

APP_PASSWORD=choose-a-password docker compose up -d --build
# app -> http://<server>:8000   (PostgreSQL + zero-shot classifier + OCR, no model downloads at runtime)

The image bakes the zero-shot DeBERTa weights, so it runs air-gapped — no calls to huggingface.co. SQLite single-container fallback: DATABASE_URL=sqlite:////app/data/contracts.db.

Defaults (the safe, clean ship config)

Setting	Default	Meaning
CLASSIFIER	auto	zero-shot DeBERTa (0.61, MIT) if weights present → rules. Clean licence.
CORE_LLM_BACKEND	auto	Ollama if reachable → rules. Extraction works with or without an LLM.
CORE_LLM_MODEL_OLLAMA	qwen2.5:7b	the extraction LLM (swap to your trained model — see below)

Plugging in the trained Ollama model (when ready)

1. Create it in Ollama (e.g. from your GGUF + Modelfile): ollama create contract-llm -f Modelfile
2. Point the app at it: CORE_LLM_MODEL_OLLAMA=contract-llm CORE_LLM_BACKEND=ollama docker compose up -d No code change — the extraction layer reads the model name from the env var.

Optional: the fine-tuned classifier (accuracy upgrade)

Default ships on zero-shot (clean licence, robust). To run the fine-tuned fusion classifier, drop the model dir in and set one env var — see docs/ACTIVATE_FINETUNED.md. Deploy on DeBERTa 0.67 (MIT), not LegalBERT (CC-BY-SA).

Verify a deployment

curl -fsS http://<server>:8000/api/health         # {"status":"ok", ...}
cd backend && ../.venv/bin/python tests/test_risk_golden.py   # golden test OK

Accuracy (for the slide)

rules 0.50 → zero-shot DeBERTa 0.61 (shipped default) → fine-tuned fusion 0.68 (measured, optional upgrade). The dynamic baseline — computed from 510 CUAD contracts — runs on top of any of them.

Status

Shippable now on zero-shot + dynamic baseline. Open/optional: trained Ollama model (hook ready above), DeBERTa-0.67 classifier weights (clean-licence accuracy upgrade), merge to main.