🧠 Fable-5 Premium V2

A rigorously cleaned, high-quality supervised fine-tuning (SFT) dataset of 100,000 agent traces, built for training tool-using models. Successor to fable-5-premium.

Priorities: Quality > Ease of Access > Quantity

📊 Dataset Overview

📦 Formats Available

This dataset is available in two formats (identical rows, identical session_id):

1. OpenAI Chat Format — Standard messages array with user/assistant/tool roles, including tool_calls and tool_call_id. Ready for Axolotl, Unsloth, TRL, and OpenAI fine-tuning API.
2. Hugging Face Agent Traces Format — Native HF Agent Traces viewable in Data Studio.

🔗 Sources

🧹 Quality Pipeline

1. Deduplication — Cross-source SHA-256 hashing on normalized messages JSON (priority: base > Crownelius > Manusagents)
2. Tool Call Validation — Every tool_calls[].id has a matching tool output, validated end-to-end
3. PII Scrubbing — Local paths, API keys (sk-*, hf_*, AKIA), emails, and IPs replaced with [REDACTED]
4. Content Filtering — Refusal patterns and out-of-range lengths removed
5. Quality Scoring — Length, structure, uniqueness, and reasoning depth; only the 0.8+ band kept

📈 Quality Distribution

All 100,000 traces fall in the 0.8–1.0 quality band (mean 0.966). Lower-scoring traces were filtered out during the quality pipeline rather than shipped.

🎯 Usage

With Hugging Face Datasets

from datasets import load_dataset

# Load OpenAI Chat format
dataset = load_dataset("saidutta69/fable-5-premium-v2", data_files="openai_chat/train.parquet", split="train")

# Load Agent Traces format
traces = load_dataset("saidutta69/fable-5-premium-v2", data_files="agent_traces/train.parquet", split="train")

With Axolotl

datasets:
  - path: saidutta69/fable-5-premium-v2
    type: chat_template
    data_files: openai_chat/train.parquet
    split: train

With Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/llama-3-8b",
    max_seq_length=4096,
)

🧰 Tool-Use Structure

• tool_calls field — Agent tool invocations with IDs, preserved per turn
• Matching tool outputs — Every tool call has its corresponding result, so models learn complete tool loops: when to call, what to pass, and how to use the result