🌊 Fathom-12K

Verified multi-turn tool-use trajectories generated from executable environments

🤗 Hugging Face Dataset · 12,745 trajectories · Apache-2.0

Code  ·  Grove-500  ·  Grove-5K

---

Fathom-12K contains 12,745 automatically verified multi-turn tool-use trajectories generated from the executable environments in Grove-500. The release is packaged as a JSONL trajectory pool.

📦 At a Glance

Field	Value
Trajectories	12,745
Source environments	500
Messages	397,183
Tool calls	105,364
Tool responses	105,364
File	fathom-12k.jsonl

🚀 Quick Start

Download the dataset with the Hugging Face CLI:

uvx hf download zhangdw/fathom-12k \
  --type dataset \
  --local-dir fathom-12k

Load the downloaded JSONL with 🤗 Datasets:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="fathom-12k/fathom-12k.jsonl",
    split="train",
)

print(dataset)
print(dataset[0]["messages"][:3])

🧱 Data Format

Each line is a JSON object with one messages array:

{
  "messages": [
    {"role": "system", "content": "You are an assistant operating in ..."},
    {"role": "user", "content": "Please help me ..."},
    {
      "role": "assistant",
      "content": "<tool_call>\n{\"name\":\"...\",\"arguments\":{...}}\n</tool_call>"
    },
    {
      "role": "user",
      "content": "<tool_response>\n{\"success\":true,\"data\":{...}}\n</tool_response>"
    }
  ]
}

Tool calls and responses are serialized inside message content using <tool_call> and <tool_response> blocks.

🎯 Intended Use

Fathom-12K is designed for:

• supervised and reinforcement-learning data preparation;
• capability-boundary and multi-turn tool-use analysis;
• conversion to model-specific function-calling formats.

📚 Citation

If you use this dataset in your research, please cite our paper published in EMNLP 2026 (Main Conference):

From Agent Skills to Executable Environments: Frontier-Targeted Training for Tool-Use LLMs

@inproceedings{zhang-etal-2026-agent-skills,
  title = "From Agent Skills to Executable Environments: Frontier-Targeted Training for Tool-Use {LLM}s",
  author = "Zhang, Dawei and Huangfu, Yubin and Yao, Kyle Yusong and
            Wei, Boyu and Han, Zongqi and Wu, Likang and
            Zhao, Hongke and He, Ming and Fan, Jianping",
  booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
  year = "2026",
  pages = "7026--7050",
  url = "https://aclanthology.org/2026.emnlp-main.305/"
}

📄 License

Fathom-12K is released under the Apache License 2.0.