Hermes Reasoning Tool Use: 51,004 Conversations for BFCL v3 Alignment

51,004 ShareGPT-formatted conversations designed to train large language models on tool-calling logic, including single-turn and multi-step reasoning. Released by @interstellarninja in late 2025, the data was generated using the Nous Research Atropos RL stack and a custom MultiTurnToolCallingEnv. The collection is specifically aligned with Berkeley Function Calling Leaderboard (BFCL) v3 evaluation scenarios.

Use Cases

• Fine-tuning models for multi-turn tool interactions using ShareGPT-formatted dialogues
• Training models to determine tool relevance in multi-step reasoning tasks
• Optimizing model performance for BFCL v3 evaluation benchmarks

Strengths

• 51,004 conversation records
• Aligned with BFCL v3 evaluation scenarios
• Covers four distinct interaction types including multi-step and relevance

Limitations

• Synthetic data generated via RL stack rather than human-tool logs
• Limited to English language text

Provenance

Source

@interstellarninja via Hugging Face

Collection Method

Synthetic generation using Nous Research Atropos RL stack and custom MultiTurnToolCallingEnv

Freshness

Last updated December 2025.

Geography

United States