saidutta69/fable-5.1-premium-v2

🧠 Fable-5.1 Premium V2 (Tool Schema)

A 9,680-trace supervised fine-tuning dataset of Fable 5.1 max-reasoning agent traces, in two tool families, each with a real tools schema column.

Built only from already published apache-2.0 traces. No distillation, no teacher calls, no new synthetic data - the work here is repair, validation, deduplication, and packaging.

Fourth entry in the Premium series, after fable-5-premium, fable-5-premium-v2 and fable-5.1-premium.

Priorities: Quality > Ease of Access > Quantity

📊 Dataset Overview

Property	Value
Total traces	9,680
fable_code	4,820 (train 4,098 / validation 361 / test 361)
claude_code	4,860 (train 4,132 / validation 364 / test 364)
Tool schemas	present on every row (tools, JSON string)
Formats	JSONL + Parquet per split
License	Apache-2.0 (inherited from sources)
Created	2026-11-01

🆕 What is new versus v1

fable-5.1-premium (v1) was 4,996 rows taken from the published 5k split only, and it did not ship a tools column.

V2 changes three things:

1. Restores tools on every row. v1 carried messages only, so a trainer had to guess the tool surface. V2 ships the schema.
2. Adds 4,860 genuinely new rows. The published 10k split contains a clean half and an anonymized half. SHA-256 comparison shows zero overlap with the 5k split. The clean half was repaired and validated; the anonymized half (5,000 rows of unknown_tool) was excluded.
3. Repairs broken tool chains. 5,875 of 423,305 tool calls (1.4%) had no matching tool response. Unmatched calls and orphan tool messages were dropped, not guessed.

🧰 Tool families

config	tools	source naming
fable_code	write_file, bash, read_file, search, list_dir	snake_case, shipped by the source
claude_code	Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch, TodoWrite, TaskCreate, TaskUpdate, TaskOutput, ToolSearch, NotebookEdit	PascalCase Claude Code surface

Important: the claude_code schemas are reconstructed, not the original system prompts. They were inferred from the observed tool calls, then every call was validated against its schema (required fields and types). 140 rows whose arguments did not fit were dropped. Treat the schemas as a faithful but derived interface.

🔗 Sources

Source	Raw	Kept	License
MoreThought Fable-5.1-Max-Reasoning-Filtered 5k	5,000	4,820	Apache-2.0
MoreThought Fable-5.1-Max-Reasoning-Filtered 10k, clean half	5,000	4,860	Apache-2.0
Same 10k, anonymized half	5,000	0	excluded (unknown_tool)
mondk Fable-5.1-Max-CLEANED	41	0	audited, no tool schemas

No other source was used.

🧹 Quality pipeline

1. Normalization - messages re-serialized canonically for hashing
2. Deduplication - SHA-256 over normalized messages plus tool names
3. Tool-chain repair - unmatched tool_calls and orphan tool messages removed
4. Schema reconstruction - the 14-tool Claude Code surface, inferred and validated
5. Argument validation - required fields and JSON types checked on every kept call
6. Quality scoring - structure, tool integrity, depth, multi-turn; only the 0.85+ band kept
7. Refusal filter - assistant refusal boilerplate removed
8. Splits - 85 / 7.5 / 7.5, fixed seed 1337

🎯 Usage

from datasets import load_dataset

fc = load_dataset("saidutta69/fable-5.1-premium-v2", "fable_code", split="train")
cc = load_dataset("saidutta69/fable-5.1-premium-v2", "claude_code", split="train")

import json
row = cc[0]
messages = json.loads(row["messages"])   # role/content/tool_calls/tool_call_id
tools    = json.loads(row["tools"])      # OpenAI-style function schemas

Schema: messages (JSON string), tools (JSON string), source, tool_family, quality_score.

❓ FAQ

What is this dataset?

9,680 Fable 5.1 max-reasoning agent traces from published Apache-2.0 sources, split into two tool families, each row carrying a JSON tools schema.

Is this newly generated data?

No. Nothing was distilled or generated here. Every row comes from an already published dataset; this release repairs, validates, deduplicates, and re-packages it.

How do I train on it?

Load either config, json.loads the messages and tools columns, and pass them to Axolotl, Unsloth, TRL, or the OpenAI fine-tuning format. Train on one config at a time: the two tool families use different names and should not be mixed in one run.

🔗 Related

• saidutta69/fable-5.1-premium - the v1, 4,996 rows, no tools column
• saidutta69/fable-5-premium - the original Premium release
• saidutta69/fable-5-premium-v2 - the 100k multi-teacher release
• RACER IS OP - Heretic Models - the models trained on this series

Responsible use

These are agent traces of a coding assistant, not a safety benchmark. The data can contain commands, file contents, and shell output from real sessions. Review before training on untrusted pipelines. Source license is Apache-2.0; keep attribution.

Made with care by RACER IS OP (saidutta69).