AAGD091/mexican-labor-llm-benchmark

Mexican Labor-Law LLM Benchmark

Dataset description

This research benchmark contains 90 Spanish-language Mexican labor-law case scenarios, reference decisions, statutory references, and associated evaluation information. It supports comparison of outcome prediction and statutory grounding across four architectures and five quantized model configurations.

The retrieval corpus used in the retained index consists of 321 SCJN resolutions and nine Mexican statutory documents (330 PDFs in total). These retrieval documents are distinct from the 90 benchmark scenarios.

Proposed release contents

benchmark/
  gold_prediction_90_compatible.json
  cases.jsonl                  # same 90 case records for the dataset viewer
evaluation/
  rubrica_validacion_contratos.json
results/
  draft_benchmark_table.csv
  draft_per_case_long.csv
  <arm>__<model>/              # checkpoint files, original folder names
  _benchmark_final/            # original aggregate and case-level CSVs
manifests/
  Verified_Retrieval_Corpus_Manifest.csv
  faiss_index_manifest.json
docs/
  PROVENANCE.md
  REPRODUCIBILITY.md

Saved filenames containing draft are retained for traceability; their status must be explicitly confirmed for the release. Raw court PDFs, statute PDFs, graph files, and serialized retrieval indexes require separate provenance and redistribution review before inclusion. Model weights and caches are excluded.

Schema

The benchmark JSON has top-level metadata and cases fields. Each case contains:

Field	Meaning
case_id	Stable identifier, PRED-001 through PRED-090
input_summary	Concise scenario description
facts	Scenario facts supplied for the task
gold_standard	Reference decision, confidence, precedent references, and key issues
expected_cel_output	Expected structured legal output, including statutory citations
benchmark_metadata	Difficulty, category, and related case attributes

The three decision classes contain 49, 26, and 15 cases, respectively. The released benchmark is an evaluation set; no independent train/validation split is established by the retained assets. In A4, the benchmark also serves as the exemplar bank, with self-exclusion by evaluated case ID.

The saved case-level results CSV contains arm, model, case_id, gold, pred, correct, and article_f1. It has 1800 rows: 90 cases for each of 20 conditions. Aggregate results include prediction, text-similarity, statutory-grounding, truncation, and extraction metrics.

Construction and validation

All 90 scenarios and outcome labels were written by two practising attorneys, each writing separate cases; the per-case assignment was not recorded. Labels are the attorneys' professional determination for the written scenarios, not recorded court rulings. Initial batch PRED-001–060; added batch PRED-061–090. No written annotation protocol or independent double annotation was used, and no agreement statistic exists. Version 1.1 removed an internal planning note that stated the added cases required double annotation; that step was not carried out (see metadata.annotation_provenance and CHANGELOG); no agreement score should be invented.

Uses and limitations

Intended for research on legal-language-model evaluation and reproducibility. The benchmark does not establish reliability for deciding real legal disputes. Its jurisdiction, domain, small sample, quantized models, unequal generation budgets, and unvalidated human–judge agreement limit generalization.

Reference decisions and citations must not be supplied to a model as evaluation inputs unless the experimental protocol explicitly calls for them. A4's reward and exemplar setup should be reported when reproducing its results.

Availability, rights, and citation

This dataset is publicly available at Hugging Face. Companion code and notebooks are publicly available at GitHub. Public access was verified without signing in on 3 October 2026. The authors approved CC BY 4.0 for their original annotations and documentation. See LICENSE-DATA.md for the scope; this does not relicense third-party material. Third-party materials retain their own terms. A complete dataset citation remains to be finalized. No archival DOI is claimed.

Versions

1.1 (2026-10-26): metadata-only update of benchmark/gold_prediction_90_compatible.json (no case record changed); in draft_benchmark_table.csv, A4 (RL+LLM) truncation is reported as not measured (empty) instead of 0, because A4 token usage was not stored. See CHANGELOG.md. 1.0 (2026-10-03): initial release.