maralzar
Switch HF Space SDK from streamlit to docker
27c67d2
|
Raw
History Blame Contribute Delete
2.52 kB
---
title: EPCC Contract Intelligence
emoji: πŸ“‘
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit
short_description: Local clause/risk classifier for EPCC contract triage
---
# EPCC Contract Intelligence β€” live demo
A leakage-safe, severity-weighted clause/risk classifier for Engineering,
Procurement, Construction & Commissioning (EPCC) contract review. Fully
local model β€” no external API calls at inference time.
- **Model**: `BAAI/bge-small-en-v1.5` sentence embeddings + logistic
regression head, trained on the public
[CUAD](https://huggingface.co/datasets/dvgodoy/CUAD_v1_Contract_Understanding_clause_classification)
dataset (13,155 attorney-labelled clauses across 509 commercial contracts,
41 clause types).
- **Held-out test** (102 contracts, β‰ˆ2,850 clauses, zero contract overlap
with training): macro-F1 0.637, critical-tier false-negative rate 0.075,
abstain at pβ‰₯0.45 β†’ coverage 0.567 / accuracy 0.903.
- **Source code, executive memo, notebook, and tests**:
[github.com/maralzar/epcc-contract-intelligence](https://github.com/maralzar/epcc-contract-intelligence)
## How to use
1. **Classify a clause** β€” paste your own or pick one of the 15 synthetic
EPCC excerpts (C1–C15). The slider tunes the abstain threshold; clauses
with confidence below it route to human review.
2. **Synthetic EPCC packet** β€” see the full pre-computed risk register for
all 15 excerpts.
3. **About / metrics** β€” held-out test metrics, the severity-tier map, and
stated limitations.
## What you're looking at
The model is a methodology proof: CUAD covers commercial / M&A contracts,
not EPCC, so the demo shows transferable behaviour. The classifier predicts
confidently on liability-cap / LD / insurance clauses (which match CUAD
labels) and **correctly abstains** on EPCC-native concepts CUAD never saw
(force majeure, site conditions, change in law, notice/waiver). The abstain
layer is doing real product work β€” that's the design point.
## Limitations
- Trained on CUAD, not on real EPCC contract data. Production needs SME-
labelled EPCC data on top of an extended taxonomy.
- Single-label per clause; real EPCC review is multi-label.
- Text-only ingestion. Real bid packages arrive as scanned PDFs; OCR is
part of the production architecture, not this demo.
- The system is **assistive**, not autonomous. Legal & commercial teams
remain final decision-makers.
License: MIT. CUAD is CC-BY-4.0 (Hendrycks et al., NeurIPS 2021).