CoolFace
Apppublic

Eva1307/legal-review-env

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes
App README

Legal Review Environment

An OpenEnv RL environment where an LLM agent reviews legal contract clauses clause-by-clause, identifies risks, and suggests edits. Models real work done by legal reviewers at startups and enterprises.

Motivation

Contract review is a high-stakes, time-consuming task. This environment trains and evaluates agents on realistic legal documents with deterministic graders — filling a genuine gap in the OpenEnv ecosystem.

Observation Space

FieldTypeDescription
task_idstringWhich task is active
contract_textstringCurrent clause title + text
clause_indexintWhich clause (0-indexed)
total_clausesintTotal clauses in contract
reviewedsofarlistHistory of previous actions + scores
step_numberintCurrent step
contextstringTask instructions

Action Space

FieldTypeValues
action_typestringflag, approve, request_edit
risk_levelstringhigh, medium, low, none
risk_categorystringindemnity, ip, termination, liability, none
suggested_editstringRequired if actiontype=requestedit
reasoningstringOne sentence explanation

Tasks

Task IDNameDifficultyDescription
task_easyClause ClassificationEasyFlag or approve clauses, identify risk level
task_mediumMulti-factor Risk AssessmentMediumAssess risk level, category, and action across a data processing agreement
task_hardRedline and EditHardPropose specific redlined edits for all risky clauses in an enterprise license

Reward Function

  • —Partial rewards at every step (not sparse)
  • —Bonuses for catching high-risk clauses with correct reasoning
  • —Penalties for approving dangerous clauses (indemnity, unlimited liability)
  • —Edit quality scored via keyword matching for task_hard

Baseline Scores (Qwen/Qwen2.5-72B-Instruct)

TaskAvg RewardSuccess
task_easy0.77true
task_medium0.84true
task_hard0.60true

Setup

bash
git clone https://huggingface.co/spaces/Eva1307/legal-review-env
cd legal-review-env
pip install -r requirements.txt
python app.py

Docker

bash
docker build -t legal-review-env .
docker run -p 7860:7860 legal-review-env

API Endpoints

  • —POST /reset?task_id=task_easy — start new episode
  • —POST /step?task_id=task_easy — submit action
  • —GET /state?task_id=task_easy — current state
  • —GET /tasks — list all tasks
  • —GET /health — health check

Inference

bash
export HF_TOKEN=your_token
export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
python inference.py