Datasetpublic

zactheaipm/agent-tool-use-synthetic

Synthetic Tool-Use Training Data for Agent Behavioral Traits Synthetic multi-turn tool-use conversations designed for mechanistic interpretability research on LLM agent behaviors. Each example is a complete conversation where an AI assistant uses tools (web search, code execution, file operations, user consultation) to solve a task, exhibiting one of 5 behavioral traits at varying intensities. Training pipeline: zactheaipm/qwenscope Traits Trait Train Eval… See the full description on the dataset page: https://huggingface.co/datasets/zactheaipm/agent-tool-use-synthetic.

sourceHugging Facemitupdated 7mo agoView on Hugging Face

0likes48downloads

Dataset Card

Synthetic Tool-Use Training Data for Agent Behavioral Traits

Synthetic multi-turn tool-use conversations designed for mechanistic interpretability research on LLM agent behaviors. Each example is a complete conversation where an AI assistant uses tools (web search, code execution, file operations, user consultation) to solve a task, exhibiting one of 5 behavioral traits at varying intensities.

Training pipeline: zactheaipm/qwenscope

Traits

Trait	Train	Eval	Description
`AUTONOMY`	1,966	~179	Acting independently vs. checking in with the user
`TOOL_USE`	1,985	~179	Frequency and appropriateness of tool calls
`PERSISTENCE`	1,953	~179	Retrying/recovering from failures vs. giving up
`RISK_CALIBRATION`	1,959	~179	Caution level when facing uncertain/risky actions
`DEFERENCE`	1,956	~179	Following user instructions literally vs. exercising judgment

Dataset Structure

—Train: 9,819 examples
—Eval: 893 examples
—Format: JSONL with OpenAI-style chat messages

Each example contains:

json

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": null, "tool_calls": [...]},
    {"role": "tool", "content": "...", "name": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "_generation_trait": "AUTONOMY",
  "_generation_model": "deepseek-chat",
  "_generation_provider": "openai"
}

Generation Details

—Generator model: DeepSeek-Chat (via OpenAI-compatible API)
—Success rate: ~98.5% (after JSON parsing cleanup)
—Scenario types: 20 diverse task categories
—Domains: 15 professional domains
—Methodology: FAST-style contrastive generation — each scenario is generated at multiple trait intensity levels

Intended Use

This dataset was created for the QwenScope project to train Sparse Autoencoders (SAEs) on Qwen 3.5-35B-A3B activations. The trained SAEs are available at zactheaipm/qwen35-a3b-saes. The contrastive trait labels enable identification of features that causally drive specific agent behaviors.

The data can also be used for:

—Fine-tuning agent behavior
—Studying tool-use patterns in LLMs
—Behavioral steering research

License

MIT