Nemotron Pretraining Legal V1: Synthetic Data for Legal LLM Capabilities

A 10-point boost in a proxy LegalBench average accuracy, from 64.6 to 74.7, was observed when this data was added to Nemotron 3 Nano pretraining. The dataset is a collection of synthetic datasets designed to improve the legal capabilities of large language models. It was created by NVIDIA for the Nemotron 3 family of models and was last updated on June 4, 2026.

Use Cases

• Pretraining legal domain language models based on synthetic legal text
• Benchmarking LLM performance on legal tasks using the proxy LegalBench metric
• Ablation studies on the impact of synthetic data on model capabilities

Strengths

• Demonstrated a 10.1-point improvement on a proxy LegalBench metric when used in pretraining
• Designed specifically for the NVIDIA Nemotron 3 family of models

Limitations

• Dataset size, row count, and file formats are unknown, limiting suitability assessment
• Column-level documentation is absent; field semantics must be inferred after download
• The description is brief; actual data quality requires manual inspection after download

Provenance

Source

NVIDIA

Collection Method

Synthetic data generation

Freshness

Last updated 2026-06-04 05:23:13