OAB-Bench

| Paper | Code |

OAB-Bench is a benchmark for evaluating Large Language Models (LLMs) on legal writing tasks from the Brazilian Bar Examination (OAB) Phase 2. The benchmark comprises 210 questions across seven areas of law from six editions (39–44) of the exam.

• Evaluates LLMs on their ability to write legal documents and answer discursive questions
• Includes comprehensive evaluation guidelines used by human examiners
• Supports two judge output formats: structured (recommended) and non-structured
• The evaluation pipeline uses LLMs as automated judges, achieving strong correlation with human scores

Note: This is the v2 release (210 questions, exams 39–44). The original release (105 questions, exams 39–41) is available at the GitHub tag v1.0.

Dataset Structure

The dataset has two configurations:

• questions (210 examples): the exam questions. Each example contains:
  • question_id: unique identifier (e.g. 39_direito_administrativo_peca_profissional)
  • category: law area for the given edition (e.g. 39_direito_administrativo)
  • statement: the question/prompt text
  • turns: conversation turns
  • values: the maximum score(s) for the item
  • system: system prompt used during generation
• guidelines (210 examples): the official evaluation guidelines used by human examiners, keyed by the same question_id.

from datasets import load_dataset

questions = load_dataset("maritaca-ai/oab-bench", "questions", split="train")
guidelines = load_dataset("maritaca-ai/oab-bench", "guidelines", split="train")

The full evaluation pipeline (model answer generation, LLM-as-judge, scoring) is available in the GitHub repository.

Results

Evaluation of 12 LLMs on OAB-Bench using GPT-5.2 as judge with structured output:

Model	Average Score	Passing Rate
Gemini-3.1-Pro	9.39	42/42 (100%)
Claude Opus 4.6	8.82	42/42 (100%)
GPT-5.2	8.66	42/42 (100%)
Claude Sonnet 4.6	8.27	42/42 (100%)
Sabiá-4	7.96	42/42 (100%)
Gemini-3.1-Flash-Lite	7.60	39/42 (93%)
Sabiazinho-4	6.91	34/42 (81%)
Sabiá-3.1	6.91	35/42 (83%)
Qwen3.5-397B	6.69	31/42 (74%)
GPT-5-Mini	6.52	31/42 (74%)
Qwen3.5-35B	6.17	24/42 (57%)
Sabiazinho-3	6.12	26/42 (62%)

Passing rate indicates the number of exams (out of 42) where the model scored ≥ 6.0. Each exam corresponds to one of the seven areas of law in a given edition.

Citation

If you find this work helpful, please cite our paper:

@inproceedings{10.1145/3769126.3769227,
  author = {Pires, Ramon and Malaquias Junior, Roseval and Nogueira, Rodrigo},
  title = {Automatic Legal Writing Evaluation of LLMs},
  year = {2026},
  isbn = {9798400719394},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3769126.3769227},
  doi = {10.1145/3769126.3769227},
  booktitle = {Proceedings of the Twentieth International Conference on Artificial Intelligence and Law},
  pages = {420--424},
  numpages = {5},
  keywords = {Open-ended Tasks, Legal Writing, Automatic Evaluation, Brazilian Bar Exam, LLM Judge, Large Language Models},
  series = {ICAIL '25}
}