Claude Fable 5: Agent Traces for AI Training

Agent Traces from the Claude Fable 5 project, authored by AyoubChLin. The dataset was last updated on June 16, 2026. It is intended for pooling together traces from various harnesses to create a meaningful training corpus.

Use Cases

• Training AI agents on tool-use patterns based on the described traces.
• Evaluating agent performance and failure modes based on traces that include limit hits and model switches.
• Benchmarking agent frameworks by converting traces to a standardized chat format using the mentioned `teich` package.
• Studying multi-turn agent interactions and decision-making processes.

Strengths

• Traces include real-world agent scenarios with events like hitting limits and model switches.
• The author provides a specific tool (`teich` package) for conversion and filtering, suggesting structured processing.
• Last updated on June 16, 2026, indicating recent activity.

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 sample data are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Collected agent traces from the Claude Fable 5 project, likely from various user harnesses.

Freshness

Last updated 2026-06-16 09:06:36