README.md

---
license: mit
language:
  - en
tags:
  - fable-5
  - claude
  - agent-traces
  - coding
  - tool-use
  - sft
  - fine-tuning
  - distillation
pretty_name: Fable-5 Premium Dataset
task_categories:
  - text-generation
  - token-classification
size_categories:
  - 10K<n<100K
---

# 🧠 Fable-5 Premium Dataset

<div align="center">
  <img src="https://res.cloudinary.com/cmazqjs6/image/upload/racer_is_op_banner_branded_pu7zud.png" alt="RACER IS OP" width="100%">
</div>

<br>

A **rigorously cleaned, high-quality** supervised fine-tuning (SFT) dataset built from Claude Fable-5 agent traces.

> **Priorities:** Quality > Ease of Access > Quantity

## 📊 Dataset Overview

| Property | Value |
|----------|-------|
| **Total Records** | 12,730 |
| **Train Split** | 5,728 (45.0%) |
| **Validation Split** | 318 (2.5%) |
| **Test Split** | 319 (2.5%) |
| **Created** | 2026-07-30 |
| **License** | MIT |

## 📦 Formats Available

This dataset is available in **two formats**:

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

## 🔗 Sources

| Source | Fable-5 Rows | Description |
|--------|-------------|-------------|

## 🧹 Quality Pipeline

1. **Deduplication** — SHA-256 content hashing across all sources (cross-source dedup)
2. **Structural Validation** — Valid message schemas, tool call IDs, proper role sequencing
3. **Content Filtering** — Remove empty/truncated responses, error-only sessions, placeholders
4. **PII Scrubbing** — Remove local paths, API keys, environment-specific data
5. **Tool Call Validation** — Ensure tool calls have matching tool responses
6. **Quality Scoring** — Multi-dimensional quality metrics

## 📈 Quality Distribution

<div align="center">
  <img src="https://huggingface.co/datasets/saidutta69/fable-5-premium/resolve/main/quality_distribution.png" alt="Quality Distribution" width="100%">
</div>

| Range | Count |
|-------|-------|
| 0.3-0.5 | 448 |
| 0.7-0.8 | 532 |
| 0.8-0.9 | 3,736 |
| 0.9-1.0 | 6,740 |

## 🎯 Usage

### With Hugging Face Datasets

```python
from datasets import load_dataset

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

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

### With Axolotl

```yaml
# axolotl config
dataset:
  - path: saidutta69/fable-5-premium
    type: chat_template
    split: train
```

### With Unsloth

```python
from unsloth import FastLanguageModel

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

## 🏗️ Chain-of-Thought (CoT)

- **`reasoning` field** — Separate field for models that support explicit thinking tokens
- **Embedded `<think>` tags** — CoT merged into assistant content for standard fine-tuning