Most modern agent datasets show long, expensive reasoning traces or only final actions. grug-think flips that: it supplies real agent trajectories rewritten so the assistant’s internal reasoning is extremely short and focused (the "grug" style) while keeping the user-facing text and tool calls unchanged. That makes it a practical resource when token cost and concise pre-tool reasoning matter.
What Sets It Apart
- Compact internal traces: median 11 words per
<think>trace (mean ~12.6, p90 = 20), explicitly optimized to teach models to reason briefly before calling tools — so models learn to be effective without long token-heavy CoT. - Fidelity to agent behavior: each example is a full system/user/assistant/tool conversation with validated tool-call JSON; only the assistant’s internal thought is rewritten into the grug format, preserving real-world action sequences.
- Broad agent work coverage: ~45% of examples are software-engineering/tooling trajectories (file edits, bash, tests), the rest cover multi-tool API interactions and single-call flows — useful for both code-focused and API-integration agents.
- Data provenance & hygiene: aggregated from multiple open-agent sources, filtered, deduplicated, and validated to avoid broken tool-call JSON and excessively long thoughts.
Who It's For
Great fit if you want to fine-tune an assistant to always emit a short, explicit internal rationale before invoking functions or tools (reducing token cost and encouraging decisive tool use). Also useful for researchers evaluating trade-offs between CoT length and action accuracy.
Look elsewhere if you need human-authored, multi-lingual chain-of-thought data, or if you require forward-looking, non-hindsight reasoning traces — the grug traces are synthetic (model-produced) and primarily English-only.
