Featherlabs/featherlabs_agentic_v1

Featherlabs Agentic v1

A curated, multi-turn conversational dataset for fine-tuning language models on agentic workflows, tool use, function calling, and chain-of-thought reasoning. Created by Featherlabs (operated by Owlkun).

This dataset is a compilation of samples from three public HuggingFace datasets, processed and unified into a consistent ShareGPT-style format.

Dataset Summary

Metric	Value
Total samples	14,676
Format	JSONL (ShareGPT-style)
Avg turns per sample	5.6
Max turns	21
File size	17 MB
Language	English

Composition

This dataset is a curated compilation of three public HuggingFace datasets:

Source Tag	Samples	Origin Dataset	Description
glaive_v2_tools	10,000	glaiveai/glaive-function-calling-v2	Function calling & tool-use conversations with system prompts, JSON tool calls, and multi-turn interactions
featherlabs_identity	2,350	Salesforce/xlam-function-calling-60k	Curated subset with identity-aware framing, establishing model persona and behavioral guidelines
opus_reasoning	2,326	distilled_corpus_400k_with_cot	Deep reasoning with <think> tags for chain-of-thought, covering math, coding, and analytical tasks

Message Statistics

Role	Count	Notes
assistant	36,512	Model responses, including tool calls and reasoning
user	28,542	User queries and follow-ups
system	10,000	System prompts with tool definitions and instructions
tool	7,470	Simulated tool/function execution results

Key Features

• 🔧 5,606 samples with tool calls — real function-calling patterns with JSON tool invocations and simulated results
• 🧠 2,326 samples with <think> reasoning — chain-of-thought traces that show internal reasoning before responding
• 📋 10,000 samples with system prompts — including detailed tool/function schemas in the system message

Format

Each sample is a JSON object with the following structure:

{
  "conversations": [
    {"from": "system", "value": "You are a helpful assistant with access to the following functions..."},
    {"from": "user", "value": "Can you search for books by J.K. Rowling?"},
    {"from": "assistant", "value": "{\"name\": \"search_book\", \"arguments\": {\"keywords\": [\"J.K. Rowling\"]}}"},
    {"from": "tool", "value": "{\"results\": [{\"title\": \"Harry Potter and the Philosopher's Stone\", ...}]}"},
    {"from": "assistant", "value": "I found several books by J.K. Rowling..."}
  ],
  "source": "glaive_v2_tools"
}

Roles

Role	Description
system	System prompt, often containing tool/function definitions in JSON schema format
user	Human user messages
assistant	Model responses — may contain tool call JSON, reasoning in <think> tags, or natural language
tool	Simulated results from tool/function execution

Use Cases

This dataset is designed for supervised fine-tuning (SFT) of language models to improve:

• Function calling — structured tool invocation with proper JSON formatting
• Agentic behavior — multi-step task execution with tool orchestration
• Chain-of-thought reasoning — internal reasoning before responding (via <think> tags)
• Identity consistency — maintaining a consistent persona across conversations
• Multi-turn coherence — following complex conversation threads with context

Training Example

from datasets import load_dataset

dataset = load_dataset("Featherlabs/featherlabs_agentic_v1", split="train")
print(f"Samples: {len(dataset)}")
print(dataset[0]["conversations"][:3])  # First 3 turns

Compatible Training Frameworks

• Unsloth — use standardize_sharegpt() to convert to chat template format
• Axolotl — use type: sharegpt in your config
• TRL / SFTTrainer — convert to messages format with role mapping: from → role, value → content

Intended Use

This dataset is intended for research and development of agentic AI assistants. It should be used as part of a larger training mix — not as a standalone fine-tuning dataset.

Recommended Mix

For best results, combine with:

• General knowledge datasets (e.g., SlimOrca, WizardLM) to prevent capability degradation
• Math & reasoning datasets (e.g., MetaMathQA, GSM8K) for quantitative reasoning
• Code generation datasets (e.g., Magicoder) for coding capabilities

Limitations

• Tool responses are simulated — they don't reflect real API outputs
• The opus_reasoning subset may contain partial problem statements
• Function schemas in system prompts vary in complexity and style
• Not suitable as a standalone training dataset — best used in a mix

Citation

@dataset{featherlabs_agentic_v1,
  title={Featherlabs Agentic v1},
  author={Owlkun},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Featherlabs/featherlabs_agentic_v1}
}

License

Apache 2.0