Agent trace archives often lose runtime tool metadata, which makes reproducing tool-enabled agent behavior hard. This dataset preserves raw, upload-first agent sessions while embedding a dataset-level tools snapshot and recovered schemas so traces remain usable for training and distillation even when individual sessions never invoked tools.
What Sets It Apart
- Preserves training-ready context: each newline-delimited JSON row is a full captured session with normalized assistant reasoning, optional assistant text, and tool-call events, plus an embedded dataset-level
toolssnapshot as a fallback. - Tool-schema recovery: native Claude Code imports and conservative name-derived MCP schemas are recovered so tool APIs, permission states, and runtime context survive conversion and training.
- Small, focused pack: 5 JSONL files with model metadata pointing to moonshotai/kimi-k3, making it lightweight for experimenting with agent fine-tuning, behavior cloning, or distillation workflows.
- Teich-native workflow: documented for use with Teich’s loading/conversion utilities (e.g.,
teich convert,load_traces), minimizing preprocessing surprises.
Who it's for and tradeoffs
Great fit if you need compact, real-session agent traces that preserve tool interfaces for training or distillation—especially for teams evaluating agent tool use, instruction-following, or behavior cloning on moonshotai/kimi-k3-style sessions. Look elsewhere if you require very large-scale corpora, fully labeled evaluation splits, or provenance for third-party production data—this pack is small (5 files) and optimized for Teich-driven workflows rather than turnkey benchmarks.
Where it fits
Use as a lightweight supplement to larger agent datasets when you want preserved tool schemas and run-recap context for debugging, fine-tuning, or curriculum distillation. It is most useful in pipelines that already use Teich and can apply the dataset-level tools snapshot during loading.
