README.md

metadata

license: mit
task_categories:
  - question-answering
  - text-classification
tags:
  - conversational
  - question-answering
  - agentic
  - multi-agent
  - tool-calling
  - function-calling
  - routing
  - agent-routing
  - agentic workflows
  - multi-agent coordination

OPENCHEF!: AGENTIC DATASET

This dataset contains simulated multi-agent sessions generated by OPENCHEF!: AGENTIC, designed for training AI agents in tool use and routing decisions. You can download the app from the .zip folder in this repo.

The OPENCLAW Agentic Dataset contains 770 full multi-agent interaction traces generated by OPENCHEF!:AGENTIC v2.0 — a small-scale simulated agent orchestration system built for realistic tool-calling, routing, and error-recovery training. Each session simulates a user prompt being handled by a team of specialized agents (Router, Email, Calendar, Code, Web, User Proxy, System, etc.) using ReAct-style reasoning, explicit tool calls, and dynamic routing with a deliberate 15% routing-error rate to force models to learn recovery patterns. Sessions average 8–15 steps, show 92.9% final success rate, and include rich metadata such as tool-use counts, routing decisions, timestamps, and generator version. The dataset is already split into train (539 ≈70%), validation (127 ≈16.5%), and test (104 ≈13.5%).

Intended use

Fine-tuning and evaluation of agentic LLMs on multi-agent coordination, function/tool calling, routing logic, error detection & recovery, ReAct-style reasoning, and realistic agent simulation. Especially valuable for studying how models handle routing mistakes, tool failures, and context management across specialized agents.

Key Stats

• Total sessions: 770
• Splits: train (539 ≈70%), validation (127 ≈16.5%), test (104 ≈13.5%)
• Success rate: 92.9%
• Average steps: varies (typically 8–15 per session)
• License: MIT

Dataset Summary

• Total Sessions: 770
• Train: 539 (70.0%)
• Validation: 127 (16.5%)
• Test: 104 (13.5%)
• Success Rate: 92.9%
• Generated: 2026-01-30T10:54:18.423Z

Supported Tasks

• Tool calling with function-calling format
• Multi-agent routing decisions
• Error recovery from routing mistakes
• Agent coordinationExample Tasks:
• Calendar: create/check events, conflict detection
• Email: search, read, forward/send
• Code: write/debug Node.js, TypeScript, Python scripts
• Web: search, summarize pages
• File/system: basic operations

Data Fields

Each example contains:

• id: Unique session identifier
• split: Dataset split (train/validation/test)
• task_description: User instruction for the session
• agents_involved: List of agent types in the session
• full_trace: Complete session trace with thoughts, actions, and observations
• final_success: Whether the session completed successfully
• steps_count: Number of turns in the session
• tool_use_count: Number of tool calls made
• routing_decisions: Number of routing decisions made
• metadata: Generation metadata

Data Splits

The dataset is split into train (70%), validation (15%), and test (15%) sets.

For Training Agentic Models

This dataset is optimized for training models on:

• Tool calling behavior
• Routing decisions between specialized agents
• Error recovery from incorrect routing
• Multi-agent coordination

Generation Details

• Generator: OPENCHEF!: AGENTIC v2.0
• Error Rate: 15% intentional routing errors
• Agent Types: Router, Email Agent, Calendar Agent, Code Agent, Web Agent, User Proxy, System
• Tool Types: Email search/send, Calendar management, Code execution, Web search, File operations
• Format: Function-calling with ReAct-style traces

Citation

If you use this dataset, please cite:

@software{openclaw_agentic_dataset,
  title = {openclaw_agentic_dataset},
  author = {webXOS},
  year = {2026},
  url = {webxos.netlify.app}
}

License

MIT License