Agent traces (conversation turns plus concrete tool calls and function schemas) are a higher-fidelity signal for teaching smaller models to behave like tool-aware assistants. This dataset pools anonymized fable-5 (Claude Code) traces in an agent-traces JSONL format so researchers and engineers can train, distill, or debug models that must reason about tool use and structured function calls.
What Sets It Apart
- Preserves tool-level structure: records include tool invocation metadata and the training-ready tool schema snapshot, which is rare in generic instruction corpora — so you can train models that generate correct tool arguments, not just free-form text.
- Teich-first compatibility: the author recommends using the teich package to parse, filter, and convert traces into OpenAI-style chat records, which simplifies downstream fine-tuning and anonymization workflows.
- Focused on Claude fable-5 (Claude Code): traces reflect one particular assistant implementation and its toolset, making the data especially relevant for distillation or behavior cloning targeted at Claude-like behavior.
Who It's For and Tradeoffs
Great fit if you want to distill tool-aware assistant behavior into smaller models, benchmark function-call generation, or experiment with agent skill orchestration using real traces. Look elsewhere if you need a large-scale benchmark with strict licensing/benchmarks guarantees: this dataset has a small download count and the repository lists no explicit license, so verify legal/compliance constraints before using it in production. Also, because traces come from a specific model and toolset, cross-model generalization should be evaluated carefully.
Where it fits: compared with generic instruction-tuning corpora, this dataset supplies structured tool-call examples and schema snapshots, making it more useful for training deterministic tool invocation. Compared with curated human-annotated eval sets, it emphasizes raw agent behavior (including model switches, limits, and harness artifacts) — valuable for engineering but requiring cleaning for some research uses.
