Tool-Reasoning-31K: 30,000+ Tool-Use Conversations with Reasoning Steps

DomofonResearch provides 30,764 unique tool-use conversations formatted in Anthropic-style ChatML. Each assistant step includes a first-person reasoning prefix before deciding to call a tool, reading a real tool result, and answering. The dataset, last updated on 2026-06-21, aggregates content from sources like xLAM, ToolACE, Glaive, Nous-Hermes, and Nvidia-When2Call.

Use Cases

• Training language models to generate reasoning traces before tool calls based on the structured conversation format.
• Evaluating model performance on multi-step tool-use scenarios described in the dataset.
• Benchmarking model ability to handle 'no tool fits' relevance scenarios mentioned in the description.
• Fine-tuning assistants for single-turn and multi-turn tool-calling interactions.

Strengths

• Contains 30,764 unique tool-use conversations.
• Includes a variety of scenarios: single-turn, multi-step, multi-turn, and 'no tool fits' cases.
• Assistant steps are prefixed with a short first-person reasoning element.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Freshness should be verified as the last update was 2026-06-21.

Provenance

Source

Derived from interstellarninja/hermes_reasoning_tool_use, which aggregates xLAM, ToolACE, Glaive, Nous-Hermes, and Nvidia-When2Call.

Collection Method

Aggregated from multiple existing datasets.

Freshness

Last updated 2026-06-21 17:51:26.