saidutta69/fable-5.1-premium

🧠 Fable-5.1 Premium

A rigorously cleaned, high-quality supervised fine-tuning (SFT) dataset of 4,996 Fable 5.1 max-reasoning agent traces, built for training tool-using and long-horizon reasoning models. Third entry in the Premium series, upholding the standards of fable-5-premium and fable-5-premium-v2.

Priorities: Quality > Ease of Access > Quantity

📊 Dataset Overview

Property	Value
Total Traces	4,996
Train Split	4,245 (85.0%)
Validation Split	374 (7.5%)
Test Split	377 (7.5%, incl. 1 held-out demo trace)
Average Quality	0.998 (0.8–1.0 band)
Total Token Count	~150,000,000
Average Trace Length	~30,000 tokens · 118–246 turns
Distilled From	Fable 5.1 (max reasoning effort)
Created	2026-09-20
License	Apache-2.0 (inherited from source)

📦 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 synthesized tool_calls and tool_call_id. Ready for Axolotl, Unsloth, TRL, and OpenAI fine-tuning API.
2. Agent Traces Format — Native system/user/assistant/tool roles with interleaved tool outputs, viewable in Data Studio.

Format	Path	Splits
OpenAI Chat	openai_chat/	train · validation · test (parquet + jsonl)
Agent Traces	agent_traces/	train · validation · test (parquet + jsonl)

Schema (both formats): messages (JSON string), model, source, session_id, quality_scores (JSON string).

🔗 Sources

Source	Raw	Kept	Description
MoreThought (Fable-5.1-Max-Reasoning-Filtered-5000x)	5,000	4,995	Fable 5.1 max-reasoning coding/reasoning agent traces
INONONO (fable-5.1-mario-trajectory)	1 session	1	Held-out demo: one-shot Super Mario Bros 1-1 remake in a single HTML file
KrazyKitty (...Filtered-1000x)	1,000	0	Verified strict subset of MoreThought via SHA-256 — collapsed
11-47 (Fable-5.1-Max-Reasoning-5K)	5,000	0	Verified duplicate of MoreThought via SHA-256 — collapsed

Deduplication audit: all 1,000 KrazyKitty rows and all 5,000 11-47 rows hash-match existing MoreThought rows (extra_from_*: 0 in build_stats.json). No unique data was discarded from forks.

🧹 Quality Pipeline

1. Deduplication — Cross-source SHA-256 hashing on normalized messages JSON; fork repos collapsed before ingest
2. Tool Call Validation — Every tool_calls[].id has a matching tool output with tool_call_id, validated end-to-end across all 4,996 traces in both formats
3. PII Scrubbing — Local paths (/home/*, /Users/*), API keys (sk-*, hf_*, AKIA), emails, and IPs replaced with [REDACTED_*]
4. Content Filtering — Refusal patterns and out-of-range lengths removed (5 traces dropped below the 0.8 quality band)
5. Quality Scoring — Length, structure, and reasoning depth composite; only the 0.8+ band kept

📈 Quality Distribution

All 4,996 traces fall in the 0.8–1.0 quality band (mean 0.998). 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-5.1-premium", data_files="openai_chat/train.parquet", split="train")

# Load Agent Traces format
traces = load_dataset("saidutta69/fable-5.1-premium", data_files="agent_traces/train.parquet", split="train")

With Axolotl

datasets:
  - path: saidutta69/fable-5.1-premium
    type: chat_template
    data_files: openai_chat/train.parquet
    split: train

⚠️ Traces average ~30k tokens (118–246 turns) — use max_seq_length >= 32k, or chunk for shorter contexts.

🧰 Tool-Use Structure

• tool_calls field — Agent tool invocations with synthesized 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
• Traces interleave file reads, writes, shell outputs, and verification runs from real agent sessions

🧾 Provenance

Built from the listed public Apache-2.0/MIT sources through the premium cleaning pipeline; full build stats shipped in build_stats.json. Companion datasets: fable-5-premium · fable-5-premium-v2.