Datasetpublic

melvindave/legal-synth-docs

Legal Synth Docs — LLM-as-Judge Dataset Training data for fine-tuning an LLM to evaluate contract clause quality across four dimensions: clarity, completeness, risk_balance, and enforceability (0–25 each, total 0–100). Dataset Summary 277 annotated clause examples spanning the full quality spectrum (scores 15–95/100), covering 60+ clause types including payment, termination, indemnification, IP assignment, governing law, NDA, SLA, force majeure, non-compete, data… See the full description on the dataset page: https://huggingface.co/datasets/melvindave/legal-synth-docs.

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes32downloads

Dataset Card

Legal Synth Docs — LLM-as-Judge Dataset

Training data for fine-tuning an LLM to evaluate contract clause quality across four dimensions: clarity, completeness, risk_balance, and enforceability (0–25 each, total 0–100).

Dataset Summary

277 annotated clause examples spanning the full quality spectrum (scores 15–95/100), covering 60+ clause types including payment, termination, indemnification, IP assignment, governing law, NDA, SLA, force majeure, non-compete, data protection, warranty, liability, franchise, employment, software licensing, M&A, real estate, and more.

Fields

Field	Type	Description
`instruction`	string	System prompt for the judge role
`input`	string	Contract clause to evaluate
`output`	string	JSON evaluation with scores, grade, reasoning, and recommendations

Scoring Rubric

Grade	Score Range
Excellent	85–100
Good	70–84
Fair	50–69
Poor	30–49
Unacceptable	0–29

See FINETUNING_GUIDE.md for fine-tuning instructions.