File size: 2,412 Bytes
b6c5070
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-nc-nd-4.0
language:
- en
pretty_name: ContractGraphEval
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- legal
- contracts
- knowledge-graph
- evaluation
- llm-as-a-judge
- uncertainty
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation.parquet
---

# ContractGraphEval

150 contract clauses, each with a reference graph, a model-predicted graph, scores from several judges, and uncertainty estimates. Benchmark of **"LLM as a Judge for Evaluating Contract Graphs"** (*Procedia Computer Science* 275, 2026).

[Paper](https://doi.org/10.1016/j.procs.2026.01.060) · [Code](https://github.com/moriyadechtiar/ensemble-graph-eval) · [Source dataset](https://huggingface.co/datasets/moriyad/contract-clause-graphs)

## Use

```python
from datasets import load_dataset

ds = load_dataset("moriyad/ContractGraphEval", split="validation")
s = ds[0]
s["clause"]["text"], s["judge_scores"]["hybrid"], s["error_prediction"]["p_error_calibrated"]
```

## Fields

| Field | Description |
|---|---|
| `clause`, `instruction` | The clause and the extraction prompt |
| `gold_graph`, `prediction` | Reference graph and the output of the GRPO-trained Llama-3.1-8B model |
| `judge_scores` | 8 reference-based scores in [0, 1] |
| `reference_free_judges` | 7 checks that need no reference graph, computed with the [`contract_linter`](https://github.com/moriyadechtiar/graph-grpo-lex) package |
| `uncertainty`, `risk_scores`, `error_prediction` | Judge disagreement, per-judge risk, and calibrated error probability |
| `evaluation` | `has_error`, `mean_score`, `structural_f1` |
| `source_id` | The same clause in the source dataset |

Error labels are derived from judge scores against the reference graph.

## Notes

- The samples are validation rows of the source dataset, from 3 contracts that were not used for training.

## Licence and citation

`cc-by-nc-nd-4.0`. Clause text comes from CUAD (The Atticus Project, CC BY 4.0).

```bibtex
@article{dechtiar2026llmjudge,
  title   = {LLM as a Judge for Evaluating Contract Graphs: Multi-Judge Benchmarking and Agentic Uncertainty-Aware Refinement},
  author  = {Dechtiar, Moriya and Katz, Daniel Martin and Jaume, Sylvain and Wang, Hongming},
  journal = {Procedia Computer Science},
  volume  = {275},
  pages   = {513--522},
  year    = {2026},
  doi     = {10.1016/j.procs.2026.01.060}
}
```