Fable-5-traces

Tasks:

Languages:

Size:

Tags:

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset viewer should be available soon. Please retry later.

Glint Research Fable 5 Agent Traces

Glint Research Dataset Card

Fable 5 Pi Agent Traces

A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.

Primary Config

pi_agent/train

Agent Trace preview enabled

4,665 Pi trace sessions 60 source sessions 3,799 tool actions 866 assistant text actions AGPL-3.0

Rows

4,665

one converted Pi trace per merged row

Tool Calls

81.44%

Bash, Edit, Read, Write, browser/preview tools

Reasoning Median

2,365

median characters in cot

Model

Fable 5

all rows report claude-fable-5

Overview

Glint-Research/Fable-5-traces preserves Fable 5 coding-agent behavior in two complementary surfaces:

Surface	Location	Purpose
Agent Trace view	`pi-traces/*.jsonl` through config `pi_agent/train`	Hugging Face Agent Traces / Data Studio inspection. Each file is a Pi-style trace containing a `session`, `model_change`, `thinking_level_change`, user context, and assistant reasoning plus text/tool output.
Merged training rows	`fable5_cot_merged.jsonl`	Flat JSONL for direct SFT, filtering, and custom conversion. Each row contains `context`, `cot`, `output_type`, `output`, `completion`, and provenance fields.
Raw logs	`/claude/`	Original session-level material retained for archival inspection and alternative converters.

The active dataset viewer is intentionally pointed at the Pi trace files so the Hub can render the examples as agent trajectories rather than as a plain JSONL table.

Dataset Anatomy

The merged row corpus was audited on June 19, 2026. Counts below are computed over fable5_cot_merged.jsonl and the generated pi-traces/ conversion.

Metric	Value
Total merged examples	4,665
Pi trace files	4,665
Unique source sessions	60
Model rows	4,665 `claude-fable-5`
Tool-use rows	3,799
Assistant text rows	866
Local-origin rows	3,712
Imported/HF-origin rows	953
Median rows per source session	38
P90 rows per source session	207
Largest source session slice	439 rows

Output Distribution

Output type	Rows	Share
`tool_use`	3,799	81.44%
`text`	866	18.56%

Tool Distribution

Tool	Rows
`Bash`	1,544
`Edit`	960
`Read`	443
`Write`	311
`PowerShell`	136
`WebSearch`	72
`mcp__Claude_Preview__preview_eval`	63
`WebFetch`	44
`TaskUpdate`	37
`ToolSearch`	35
`TaskCreate`	26
`mcp__Claude_Preview__preview_screenshot`	24
`ScheduleWakeup`	23
Other tools	81

Source Mix

Source root	Rows
`-home-lane-MythosMini`	2,024
imported HF slice / other	953
`-home-lane-GR`	447
`-home-lane`	425
`-home-lane-AIArchives`	316
`-home-lane-rblx`	215
`-home-lane-Blindbot-hf-space`	87
`-home-lane-AOTpy`	60
`-home-lane-Blindbot`	58
`-home-lane-log`	52
`-home-lane-letsclaudething`	28

Text Length Profile

Field	Median chars	P90 chars	P95 chars	Max chars	Mean chars
`context`	7,022	7,022	7,022	7,022	6,593.2
`cot`	2,365	4,186	5,274	9,145	2,669.4
`completion`	2,726	6,166	8,848	73,607	3,754.9

The fixed upper bound in many context rows reflects the merged-file truncation policy. Use /claude/ if you need to inspect fuller raw session logs.

Pi Agent Trace Mapping

Each merged row is converted into a minimal Pi-compatible trace:

Event	Meaning
`session`	Synthetic stable UUID derived from row UID; `cwd` normalized to `/workspace`.
`model_change`	Records `claude-fable-5`.
`thinking_level_change`	Set to `high` for trace viewer grouping.
`message` user	The merged row `context`, preserving the prompt and preceding tool/result transcript.
`message` assistant	A `thinking` item from `cot`, then either a `text` item or a `toolCall` item derived from `output`.

This mapping is designed for visualization and distillation. It does not claim that each converted file was originally a standalone Pi run; it is a faithful row-level projection of the merged Fable 5 trace data into the Hub-supported Agent Traces format.

Fields in `fable5_cot_merged.jsonl`

Field	Type	Description
`uid`	string	Stable row identifier, usually `session_id#index`.
`source_file`	string	Original raw trace source path.
`session`	string	Source session identifier.
`model`	string	Captured model name; all audited rows are `claude-fable-5`.
`context`	string	Prompt and prior transcript context used for the row.
`cot`	string	Reasoning trace captured for the assistant action.
`output_type`	string	`tool_use` or `text`.
`output`	object	Tool name and arguments for tool rows, or text payload for assistant message rows.
`completion`	string	Full serialized completion-style representation containing reasoning and output.
`origin`	string	Provenance bucket, currently `local` or `hf`.

Loading

View Agent Traces

Open the Hub viewer at:

https://huggingface.co/datasets/Glint-Research/Fable-5-traces/viewer/pi_agent/train

Stream the Pi Agent Trace Projection

from datasets import load_dataset

ds = load_dataset(
    "Glint-Research/Fable-5-traces",
    "pi_agent",
    split="train",
    streaming=True,
)

row = next(iter(ds))
print(row["harness"])
print(row["session_id"])
print(row["messages"][-1])
print(row["file_path"])

Load the Flat Merged JSONL

from datasets import load_dataset

merged = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/Glint-Research/Fable-5-traces/resolve/main/fable5_cot_merged.jsonl",
    split="train",
)

print(merged.column_names)
print(merged[0]["output_type"])

Intended Uses

SFT and distillation research for coding-agent reasoning plus action prediction.
Tool-call policy modeling over realistic shell, file-editing, preview, web, and task-management actions.
Agent trace visualization and qualitative inspection in Hugging Face Data Studio.
Debugging row-level conversions between flat training JSONL and event-oriented agent traces.
Studying how long-form reasoning, terminal observations, and concrete tool invocations interleave in coding sessions.

Caveats and Responsible Use

The dataset contains coding-agent transcripts, terminal outputs, local file paths, and generated work logs. Treat it as agent telemetry, not as sanitized benchmark data.
Many merged context values are intentionally truncated; use raw /claude/ files when session-level continuity matters.
Tool outputs and command logs are not guaranteed to be executable in your environment.
This is not a hidden-eval dataset and should not be used as a claim of model capability by itself.
The active card is optimized for Agent Trace inspection. The flat JSONL remains available for training workflows that prefer row-level examples.
License: AGPL-3.0. Check compatibility before using this data in commercial or closed-source training pipelines.

Provenance

This dataset was assembled from Fable 5 agent traces collected by Glint Research and companion trace material contributed through the TeichAI ecosystem. The Pi Agent trace projection was generated from fable5_cot_merged.jsonl so the corpus can be inspected through the Hugging Face Agent Traces viewer.