---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- agentic
- tool-calling
- openclaw
size_categories:
- 10K<n<100K
---

# ClawGym-Filtered

Filtered + reformatted OpenClaw agentic trajectories from
[RUC-AIBOX/ClawGym-Trajectory](https://huggingface.co/datasets/RUC-AIBOX/ClawGym-Trajectory).

- 24,497 raw → **15,777** rows (64.4% retention)
- **342M tokens** (Qwen3-4B tokenizer)
- Generators: MiniMax-M2.5, GLM-5.1 / Scaffold: OpenClaw (19 tools)

Assistant CoT is moved from `content` to `reasoning_content` for every
tool-call turn; only the final no-tool-call turn keeps user-visible `content`.

## Files

| Path | Format |
|---|---|
| `clawgym.jsonl` | raw final JSONL |
| `slime/clawgym.parquet` | slime 6-col, JSON-string messages/tools |
| `torchtitan/clawgym.parquet` | torchtitan 4-col, native `list<struct>` context |
| `torchtitan_sft/` | Arrow dataset (3 shards, 256K max-token cap) |