LegalBench
Homepage: https://hazyresearch.stanford.edu/legalbench/ LegalBench is a collaboratively built benchmark of legal reasoning tasks.
How to run
$ npm i -g lettokenburn-cli$ ltb bench pull legalbench --n 50$ ltb bench run ./legalbench -a claude-code -m anthropic/claude-opus-4-8$ ltb bench push
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
3 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-4 | 82.9 | N/A | Independent | medium | N/A | paperswithcode.com |
| GPT-3.5 Turbo | 60.9 | N/A | Independent | medium | N/A | paperswithcode.com |
| Claude-1 | 58.1 | N/A | Independent | medium | N/A | paperswithcode.com |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Homepage: https://hazyresearch.stanford.edu/legalbench/ LegalBench is a collaboratively built benchmark of legal reasoning tasks. The full benchmark contains 162 tasks; this implementation covers the five-task subset used by HELM (aggregated by the legalbench group) plus the complete 14-task Contract NLI suite (aggregated by the legalbench_contract_nli group), a coherent single-corpus family of Yes/No entailment questions over non-disclosure agreements. Prompts (instructions, field ordering, and output nouns) follow the official HazyResearch/legalbench prompt files — the HELM prompt settings for the HELM subset, and the authors' per-task base_prompt.txt instructions for Contract NLI — and data is loaded from the `nguha/legalbench` Hub dataset. Tasks are evaluated zero-/few-shot with generate_until and case-/punctuation-insensitive exact match; the train split provides the few-shot examples.
Available in EleutherAI lm-evaluation-harness as task family legalbench (19 tasks).