Llama 4 Maverick Distilled: 5,000 Synthetic Reasoning Traces

5,000 high-quality reasoning traces distilled to mirror the step-by-step thinking of Llama 4 Maverick class models. The dataset was synthetically generated by Meta AI and published by WithinUsAI on May 24, 2026. It is designed for training student language models to reproduce a specific chain-of-thought pattern.

Use Cases

• Train student language models to emulate chain-of-thought reasoning based on the described reasoning traces.
• Benchmark the reasoning performance of language models against a distilled reference style.
• Study the structure and patterns of synthetic reasoning generated by advanced language models.

Strengths

• Contains 5,000 explicit reasoning traces.
• Designed to capture a specific chain-of-thought pattern characteristic of a named model class.
• Sourced from synthetic generation by Meta AI, suggesting a controlled creation process.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale metrics like file size and format are unknown.
• Data may reflect the synthetic generation bias inherent to its source model.

Provenance

Source

Meta AI

Collection Method

Synthetic generation

Freshness

Last updated 2026-05-25 05:02:37.