The MFV Legal Tranche is a rights-cleared, provenance-verified legal reasoning dataset: 767 contribution cards, 119 packages, and 291 correction events across 2,389 files, derived from California civil litigation that settled favorably against a top-20 defense firm. Licensed for AI training, fine-tuning, and evaluation. A free sample is available on Hugging Face and AWS Data Exchange; the full corpus is licensed from $299/month.
| Metric | Value |
|---|---|
| Packages | 119 |
| Contribution cards | 767 |
| Correction events | 291 |
| Files | 2,389 |
| Legal failure types | 10 |
| Craft dimensions | 8 |
| Lifecycle stages | 7 |
| Severity levels | 4 |
| Validated quality lift | 20% |
| Validation method | Two independent AI judges |
Frontier-model output augmented with the dataset's retrieval library and calibrated system prompts, scored against baseline output on the 8-dimension rubric by two independent AI judges from different model families (Claude Sonnet 4.6 and GPT-4o). Overall quality lift: 20% under both judges.
| Dimension | Lift (Sonnet 4.6 judge) | Lift (GPT-4o judge) | Score reached |
|---|---|---|---|
| Jurisdictional specificity | +36% | +31% | 4.97/5 |
| Citation accuracy | +22% | +29% | 4.70/5 |
| Evidence integration | +20% | +27% | 4.20/5 |
| Argument persuasiveness | — | — | 4.67/5 under both judges |
Floor and ceiling: under Sonnet 4.6 (the stricter judge), 4.18/5 overall with 7 of 8 dimensions at strong professional quality (4.0+); under GPT-4o, 4.72/5 overall with all 8 dimensions above 4.0 and 7 of 8 at near-exceptional quality (4.5+).
Single author, 100% owned, zero scraped content, no third-party rights. Every document originates from two years of pro se California civil litigation conducted by the author against a top-20 defense firm. The origin of the material is documented in a public court record: Superior Court of California, County of Los Angeles, Case No. 25STCV14949. Post-litigation-era buyers price licensing risk into every training-data purchase — a single-author corpus with a verifiable public-record origin removes that risk entirely: there is no scraped content to trace, no third-party rights to clear, and no ambiguity about who owns the data.
The corpus itself is sanitized for public sale: party names and personally identifying details are redacted throughout, so the dataset ships clean for training use while the public court record independently verifies its origin.
Fine-tuning legal AI models. The tranche captures how legal reasoning actually unfolds in live litigation — claims constructed, opposing arguments answered, errors caught and corrected — rather than the static case-law text most legal corpora contain. Fine-tuning on reasoning traces with documented corrections teaches a model the process of legal argument, not just its vocabulary.
RAG evaluation. The 767 contribution cards are self-contained, human-authored units of legal knowledge with known relationships to the 119 packages they support. That structure makes them natural ground truth for evaluating whether a retrieval pipeline surfaces the right authority for a given legal question — a labeled retrieval corpus that did not come from synthetic generation.
Legal-reasoning benchmarks. The 291 correction events are before-and-after pairs: a flawed legal work product and its repaired version. Correction pairs are benchmark material — they let you test whether a model can spot the same defects a human litigator had to find and fix under adversarial pressure.
Also from Meta-Flywheel Ventures: the Professional Screenwriting AI Evaluation & Correction Dataset — a second rights-cleared judgment dataset from the same documented pipeline.