RiskLoop Representative Model Checkpoints
This repository hosts the reproduced representative model checkpoints for RiskLoop, an NLP system for automated contract risk detection based on Legal-BERT fine-tuning on the CUAD (Contract Understanding Atticus Dataset) benchmark.
Provenance & Reproduction Note: These checkpoints are reproduced representative checkpoints created under strictly controlled, frozen training conditions (Condition A) and evaluated in RiskLoop's frozen Phase 6 test evaluation. They are NOT the original historical official checkpoints from earlier development phases, as those original historical binaries were lost. These reproduced models represent the exact frozen representative checkpoints used to generate the Phase 6 decision analysis and final test metrics.
1. Model Architecture & Training Details
- Backbone Model:
nlpaueb/legal-bert-base-uncased - Model Type: Single-task span classification models (QA-style start/end logit heads over transformer sequence outputs).
- Training Strategy: Condition A (single-task fine-tuning with fixed seeds, 4 epochs, max sequence length 512, document stride 256).
2. Model Checkpoint Registry & Integrity Hashes
The repository contains three frozen model binary files (.pt PyTorch state dictionaries):
| Checkpoint Filename | Target Contract Task | Experimental Condition | Random Seed | File Size (Bytes) | SHA-256 Hash |
|---|---|---|---|---|---|
run_03_best_model.pt |
Cap On Liability | Condition A | 44 |
438,001,498 | 8689ef00a0718713f4ae8bdf8b48d3441b5d9a2ccd13ae322185862551c6f674 |
run_04_best_model.pt |
Anti-Assignment | Condition A | 42 |
438,001,498 | ea6a75254798342e0cc3925dd8adca1ba9b5bb60c6a631ddf14057748c01947c |
run_07_best_model.pt |
Termination For Convenience | Condition A | 42 |
438,001,498 | a8a28902ef3dbc28e706135e8599f01cfa8c0e9c4c05cc081818e3027deda119 |
3. Frozen Phase 6 Test Evaluation Metrics
The reproduced representative checkpoints achieved the following metrics on the held-out frozen CUAD test evaluation set (Phase 6):
| Task Name | Representative Checkpoint | Test F1 Score | Test ROC-AUC | Test PR-AUC |
|---|---|---|---|---|
| Cap On Liability | run_03_best_model.pt |
0.7727273 |
0.9960492 |
0.9147721 |
| Anti-Assignment | run_04_best_model.pt |
0.8083624 |
0.9944473 |
0.9146714 |
| Termination For Convenience | run_07_best_model.pt |
0.7483871 |
0.9939074 |
0.7926948 |
4. Inference Score Interpretation
In RiskLoop's inference engine and Streamlit demo interface:
- Output scores are uncalibrated raw logit deltas ($\text{logit}_1 - \text{logit}_0$).
- High positive scores indicate strong model activation for clause presence.
- Scores are not calibrated probabilities or confidence percentages.
5. Usage & Legal Disclaimer
- Intended Use: Portfolio evaluation, academic research, and interactive open-source demonstration of contract risk detection.
- Legal Disclaimer: This software and model predictions do NOT constitute legal advice, formal contract audit, or legal guarantee. Users should consult qualified legal professionals for actual contract review.