docs/ACTIVATE_FINETUNED.md

Activating the fine-tuned classifier (DeBERTa 0.67 / LegalBERT 0.70)

This branch (dynamic-finetuned) = the dynamic baseline + all features + the fine-tuned-classifier slot wired in. It runs end-to-end today on zero-shot DeBERTa (0.61, MIT). Drop in trained weights to upgrade — no code change.

Right now (no weights) — fully working

cd backend && CLASSIFIER=zeroshot CORE_LLM_BACKEND=ollama ../.venv/bin/uvicorn app.main:app --port 8000

CLASSIFIER=fusion also works and gracefully falls back to zero-shot until weights are present.

To activate the fine-tuned model

1. Get the trained model folder (gitignored — shipped separately):
   • deberta-sent-cuad/ → DeBERTa-v3, 0.67 macro-F1, MIT licence ← use this for deployment
   • legalbert-sent-cuad/ → LegalBERT, 0.70, but CC-BY-SA (avoid shipping)
2. Put it at backend/finetune/<name>/.
3. Run with the fusion classifier pointed at it:

cd backend
CLASSIFIER=fusion \
LEGALBERT_MODEL_DIR=finetune/deberta-sent-cuad \
CORE_LLM_BACKEND=ollama \
../.venv/bin/uvicorn app.main:app --port 8000

The fine-tuned model now owns its 12 CUAD categories; zero-shot DeBERTa fills the rest; the dynamic baseline runs downstream. Verify with:

cd backend && CLASSIFIER=fusion LEGALBERT_MODEL_DIR=finetune/deberta-sent-cuad \
  ../.venv/bin/python -m eval.run_eval --classifier legalbert --limit 50

To train the weights (if you don't have them)

See docs/COLAB_FINETUNE.md (GPU, ~15 min) — train with --model microsoft/deberta-v3-base for the MIT/0.67 model.

Why DeBERTa for deployment

The product ships on-prem (weights go to the client). LegalBERT is CC-BY-SA (share-alike); DeBERTa is MIT — clean to sell. The ~3-point F1 gap (0.70→0.67) is the price of a license you don't have to explain. See docs/DECISIONS.md.