LIMI: Less is More for Agency Language Model Training Data

LIMI is a dataset for training language models in tool use and agency, released by GAIR. The release includes training scripts for Qwen3 dense models of varying sizes. Its structure includes defined message schemas, tool definitions, and tool-call formats.

Use Cases

• Fine-tuning language models for tool-calling based on defined tool schemas.
• Training agentic models to follow instructions using structured message formats.
• Benchmarking model performance on tasks requiring external tool integration.

Strengths

• Includes training scripts for Qwen3 models (4B/8B/32B).
• Defines a structured schema for messages and tool calls.
• Released by the research organization GAIR.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license are unknown, which may limit suitability assessment.

Provenance

Source

GAIR

Freshness

Last updated 2025-10-09 03:20:49; freshness should be verified.