Why this matters
Agent trace datasets capture real multi-step decision and tool-use behavior that supervised fine-tuning or distillation can turn into reliable action policies. This collection bundles 244 Cursor sessions driven by Claude Fable‑5 at high effort levels, preserving assistant turns, tool calls, and terminal-style outputs so you can train or analyze agentic behaviors without reconstructing execution logs.
What Sets It Apart
- Raw agent trajectories captured from Cursor: each session records granular assistant/tool interactions and results rather than only final replies, making it suitable for next-step prediction and tool‑call policy learning.
- Teich-compatible formatting: prepared to plug into the Teich conversion/training pipeline, reducing preprocessing friction for distillation and SFT workflows.
- High-effort, behavior-rich examples: sessions include High/xHigh/Max effort runs intended to expose planning, multi-step reasoning, and repeated tool-use patterns.
- Very long rows: some rows exceed one million characters (full transcripts and tool outputs), so the dataset explicitly warns and expects you to apply a tokenizer-aware oversize policy before training.
Who it's for — and tradeoffs
Great fit if you want to: train or distill agentic behaviors (tool selection, stepwise coding/debug loops), evaluate policy prediction for coding agents, or analyze real assistant execution traces. Look elsewhere if you need large-scale, diverse user-chat datasets (this corpus is focused on Cursor-based agent traces and is relatively small by raw row count). The long-row nature gives rich context for verification-driven training but increases preprocessing and memory demands; plan tokenization, chunking, or curriculum strategies accordingly.
Where it fits
Use this dataset as a high-signal supplement when building or fine-tuning coding/agent models, modeling tool-call policies, or studying agent planning. For broad conversational pretraining or massive multilingual coverage, pair it with larger, more diverse corpora.
