Nathan9/dump/Datasets - Storage Bucket

Complete FABLE.5 Traces 2M

Provenance-cleaned FABLE.5 / Claude corpus — trimmed to content-verified traces only.

Dataset Viewer | Parquet

This dataset is a post-closure compilation of FABLE.5 / Claude trace datasets found on Hugging Face after the closure of Fable and Mythos. It is deduplicated at the normalized-row level and keeps row-level provenance through first_source_dataset, first_source_config, first_source_split, and first_source_row_index. A provenance pass has since removed material that was matched only by dataset name/keyword and was not genuine FABLE.5 / Claude data — see What Changed. This is the Claude-model half of a pair; its OpenAI-model sibling is the GPT-5.6 Sol · Terra · Luna Library.

// what this is

This is a maintained library — a community mirror of publicly-available Fable-5 / Claude datasets on Hugging Face, aggregated, cleaned, and content-verified with per-row source attribution (first_source_dataset). It is not Crownelius' own data. It exists to give Fable-5 traces one clean, honest home, with full credit to each original author — see the source citations below.

Verified Corpus Snapshot

56,700
verified rows

2,006,487
original combined size

~98%
of original removed

⚠ Correction — provenance cleanup

An earlier version advertised ~2,006,487 "FABLE.5 traces." On inspection, ~98% was not FABLE.5 / Claude data — it had been matched by dataset name/keyword ("fable", "mythos", "claude") rather than verified by row content. That contamination has been removed and the corpus rebuilt from content-verified rows only.

What Changed

A provenance and quality pass rebuilt the corpus so that every row is content-verified as genuine FABLE.5 / Claude model-interaction data (Anthropic toolu_… tool-use IDs, Claude Code harness markers, /model → Fable 5, or model: claude-fable-5 tags) — judged on row content, not dataset name. Each source verdict was adversarially re-checked.

Original combined size: 2,006,487 rows → 56,700 verified rows. That is 40,421 verified rows kept from the original (~2%) plus 16,279 content-verified rows added from 9 sources.

Removed — matched by keyword/name, not genuine FABLE.5 / Claude

Rows	Source	What it actually was
1,100,000	attentionAllYouNeed/Vibe-Coding-Claude-Fable-5	Procedurally-generated template filler (a few stems × a "Scenario N" counter); no trace structure. "Claude"/"Fable" only in the name.
775,175 of 776,496	Poumrm/Mythos-5-and-Fabel-5-Class-Model-Outputs	Generic synthetic SFT ("Class Model Outputs"). Only ~0.16% were real traces — those 1,321 were salvaged; the rest removed.
40,215	BerkayBB/MythosAi	Turkish news-summarization corpus. Matched on "mythos".
25,000	ansulev/claude_mythos_distilled_25k	Self-declared synthetic imitation — its own card states "Not real Claude/Mythos outputs."
7,635	desiree/mythos-coder-data	Procedurally slot-filled synthetic coding exercises.
7,090	umm-maybe/wikisource-cthulhu-mythos	Public-domain Lovecraft / Cthulhu-mythos fiction. Matched on "mythos".
951	Glint-Research/mythosmini	Synthetic tool-use traces of unknown provenance (upstream now private).
551	ox-ox/mythos-character-distillation	Majority Gemini/manual-generated + short-story fiction.
347	ronniealfaro/mythos	Spanish mythology/philosophy essays. Matched on "mythos".
192	HelioAI/Fable-5-Distill-Reasoning-462x	CoT distilled from an unspecified "Mythos V2" teacher — not verifiably Anthropic.

A row-level validity filter was also applied to the kept sources: only QA, chain-of-thought, and agent/tool-use rows with a real question and a substantive response are retained. Dropped: control/envelope events (queue-operation, mode, permission-mode, session pointers, UI/paste metadata), model-fallback markers, empty/null content, placeholder stubs (e.g. ***…***), author-flagged decoy rows, questions with no answer, and unstructured/whole-document dumps.

Added — content-verified and deduplicated by sha256(row_json) (9 sources, 16,279 rows)

Rows	Source	Notes
6,049	greghavens/fable-5-coding-and-debugging-traces	Verifier-backed Fable-5 coding & debugging traces (CC-BY-4.0).
5,060	lordx64/fable-tool-use-sft	FABLE.5 tool-use pairs (<tool_use … id="toolu_…"> XML format).
4,460	Glint-Research/Fable-5-traces	Canonical FABLE.5 trace hub (messages/tools/trace).
347	armand0e/Fable-5-Chat	FABLE.5 chat Q&A pairs.
174	licongxu/fable5-flamingo-research-trace	Claude Code research-session trace events.
107	cfahlgren1/Fable-5-traces	Curated Claude Code sessions.
63	suayptalha/fable-5-claude-code	Claude Code coding sessions.
18	AlinCiocan/fable-5-claude-code-traces	Claude Code sessions.
1	victor/claude-fable-worldcup-2026-session	Single long Claude Code session.

Verified-but-excluded this round: choucsan/mimo-claude-code-traces-1k (MiniMax MiMo model, mislabeled claude-fable-5), lordx64/fable-sft-combined-v2 (derivative re-combine of already-included data), Sexyeinstein/Agent (redundant raw dumps of already-included sessions), hotdogs/uka-fable-reasoning (gated — content not verifiable).

Files

File	Rows	Purpose
data/train.parquet	2,006,487 → 56,700	Canonical, viewer-friendly Parquet split.

The previous raw/fable5_mythos_dedup.jsonl.gz mirror was removed — it held the pre-cleanup contaminated export.

Provenance & honesty

Retained rows are content-verified and attributed to FABLE.5 / Claude by their upstream sources. This provenance is source-asserted and content-verified — not cryptographically certified by Anthropic, and this dataset does not claim Anthropic endorsement. There is no method to prove text was emitted by a specific model; the cleanup removes keyword/name-matched contamination and keeps genuine QA / CoT / agent-trace data.

Schema

Column	Type	Description
row_hash	string	SHA256 of the normalized original row JSON.
first_source_dataset	string	First dataset where this canonical row was observed.
first_source_config	string	First observed config or loader surface.
first_source_split	string	First observed split, file, or local extraction label.
first_source_row_index	int64	Row index within the first observed source surface.
seen_count	int64	Number of times this canonical row was seen during the pass.
row_json	string	Preserved original row as canonical JSON. Parse this for source-specific fields.

Source Citations

Every row can be traced back to the dataset where that canonical row was first observed, via the first_source_* fields.

Retained sources: 17 Hugging Face datasets (56,700 rows)

Rows	Source Dataset	Link
16,979	1EYE4ALL/Fable-5-traces	dataset
11,947	armand0e/claude-fable-5-claude-code	dataset
6,049	greghavens/fable-5-coding-and-debugging-traces	dataset
5,060	lordx64/fable-tool-use-sft	dataset
4,460	Glint-Research/Fable-5-traces	dataset
4,399	Swarm-AI-Research/fable5-traces-sft	dataset
4,022	lordx64/agentic-distill-fable-5-sft	dataset
1,321	Poumrm/Mythos-5-and-Fabel-5-Class-Model-Outputs (salvaged)	dataset
815	victor/fable-5-boeing-747-trace	dataset
591	PawanKrd/claude-fable-5-code	dataset
347	TheFusionCube/Fable-5-CoT-Traces	dataset
347	armand0e/Fable-5-Chat	dataset
174	licongxu/fable5-flamingo-research-trace	dataset
107	cfahlgren1/Fable-5-traces	dataset
63	suayptalha/fable-5-claude-code	dataset
18	AlinCiocan/fable-5-claude-code-traces	dataset
1	victor/claude-fable-worldcup-2026-session	dataset

Removed sources: 10 datasets (keyword/name matches, not genuine FABLE.5 / Claude)

Rows removed	Source Dataset	Link
1,100,000	attentionAllYouNeed/Vibe-Coding-Claude-Fable-5	dataset
775,175	Poumrm/Mythos-5-and-Fabel-5-Class-Model-Outputs (non-trace bulk)	dataset
40,215	BerkayBB/MythosAi	dataset
25,000	ansulev/claude_mythos_distilled_25k	dataset
7,635	desiree/mythos-coder-data	dataset
7,090	umm-maybe/wikisource-cthulhu-mythos	dataset
951	Glint-Research/mythosmini	dataset
551	ox-ox/mythos-character-distillation	dataset
347	ronniealfaro/mythos	dataset
192	HelioAI/Fable-5-Distill-Reasoning-462x	dataset

Loading

from datasets import load_dataset

ds = load_dataset("Crownelius/Complete-FABLE.5-traces-2M", split="train")