Spaces:
Running
Running
File size: 7,057 Bytes
04b4cff | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | {
"benchmark_card": {
"benchmark_details": {
"name": "LEGALBENCH",
"overview": "LEGALBENCH is a benchmark designed to measure the legal reasoning capabilities of large language models. It comprises 162 tasks collaboratively constructed and hand-crafted by legal professionals, covering six distinct types of legal reasoning.",
"data_type": "text",
"domains": [
"legal",
"law",
"finance"
],
"languages": [
"English"
],
"similar_benchmarks": [
"GLUE",
"HELM",
"BigBench",
"RAFT"
],
"resources": [
"https://arxiv.org/abs/2308.11462",
"https://huggingface.co/datasets/nguha/legalbench",
"https://storage.googleapis.com/crfm-helm-public/lite/benchmark_output/releases/v1.13.0/groups/core_scenarios.json"
]
},
"purpose_and_intended_users": {
"goal": "To enable greater study of what types of legal reasoning large language models (LLMs) can perform.",
"audience": [
"Practitioners (to integrate LLMs into workflows)",
"Legal academics",
"Computer scientists"
],
"tasks": [
"Text classification",
"Question answering",
"Text generation",
"Rule-application tasks"
],
"limitations": "The tasks do not generalize to all legal reasoning tasks or all types of legal documents, offering only a preliminary understanding of LLM performance.",
"out_of_scope_uses": [
"Predicting the legality of real-world events",
"Predicting the outcome of lawsuits",
"Providing legal advice"
]
},
"data": {
"source": "Data is drawn from three categories: existing publicly available datasets and corpora (some reformatted), datasets previously created by legal professionals but not released, and tasks developed specifically for LegalBench. The tasks originate from 36 distinct corpora.",
"size": "The benchmark comprises 162 tasks. The distribution of tasks by sample count is: 28 tasks have 50-100 samples, 97 have 100-500 samples, 29 have 500-2000 samples, and 8 have 2000+ samples. The overall dataset falls into the size category of 10,000 to 100,000 examples.",
"format": "Examples are presented in a structured format with fields such as 'Task name', 'Question', 'Options', and 'Answer'.",
"annotation": "Annotation procedures are task-dependent. For certain tasks, each data point was manually validated by a law-trained expert. Detailed annotation methodology for each task is documented in a separate section of the paper."
},
"methodology": {
"methods": [
"Models are evaluated in a few-shot setting. Train splits consist of a small random sample of between 2 to 8 instances to capture a true few-shot learning scenario.",
"For rule-application tasks, a law-trained expert manually validates each model generation."
],
"metrics": [
"LegalBench",
"Correctness",
"Analysis"
],
"calculation": "The primary benchmark metric is Exact Match (EM) on LegalBench. For rule-application tasks, two separate metrics are computed: 'correctness' (the proportion of generations without errors) and 'analysis'.",
"interpretation": "Higher scores on the LegalBench metric indicate better performance. The metric is continuous and lower scores are not better.",
"baseline_results": "The original paper evaluated 20 LLMs from 11 different families but did not provide specific scores. In a separate evaluation suite, the Yi 34B model achieved a score of 0.618.",
"validation": "For rule-application tasks, a law-trained expert manually validated each model generation. For datasets reused or adapted from other sources, the original data sheets document any redactions or missing data."
},
"ethical_and_legal_considerations": {
"privacy_and_anonymity": "For tasks that reuse or adapt existing datasets, the benchmark refers to the original data sheets for details on any data redactions or missing information.",
"data_licensing": "other",
"consent_procedures": "Not specified.",
"compliance_with_regulations": "The benchmark includes a section for each task intended to provide information relevant to ethical review processes, but specific details are not provided in the available facts."
},
"possible_risks": [
{
"category": "Over- or under-reliance",
"description": [
"In AI-assisted decision-making tasks, reliance measures how much a person trusts (and potentially acts on) a model's output. Over-reliance occurs when a person puts too much trust in a model, accepting a model's output when the model's output is likely incorrect. Under-reliance is the opposite, where the person doesn't trust the model but should."
],
"url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/over-or-under-reliance.html"
},
{
"category": "Unrepresentative data",
"description": [
"Unrepresentative data occurs when the training or fine-tuning data is not sufficiently representative of the underlying population or does not measure the phenomenon of interest. Synthetic data might not fully capture the complexity and nuances of real-world data. Causes include possible limitations in the seed data quality, biases in generation methods, or inadequate domain knowledge. Thus, AI models might struggle to generalize effectively to real-world scenarios."
],
"url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/unrepresentative-data.html"
},
{
"category": "Data bias",
"description": [
"Historical and societal biases might be present in data that are used to train and fine-tune models. Biases can also be inherited from seed data or exacerbated by synthetic data generation methods."
],
"url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/data-bias.html"
},
{
"category": "Lack of data transparency",
"description": [
"Lack of data transparency might be due to insufficient documentation of training or tuning dataset details, including synthetic data generation.\u00a0"
],
"url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/lack-of-data-transparency.html"
},
{
"category": "Improper usage",
"description": [
"Improper usage occurs when a model is used for a purpose that it was not originally designed for."
],
"url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/improper-usage.html"
}
],
"flagged_fields": {},
"missing_fields": [
"ethical_and_legal_considerations.consent_procedures"
],
"card_info": {
"created_at": "2026-03-17T12:59:10.203815",
"llm": "deepseek-ai/DeepSeek-V3.2"
}
}
} |