Overview

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

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.

Output Distribution

Tool Distribution

Source Mix

Text Length Profile

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:

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

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.

Related Links

• Hugging Face Agent Traces documentation: https://huggingface.co/docs/hub/agent-traces
• Pretty external viewer: https://tracehouse.ai/d/glint-research-fable-5-traces-0d89ee?t=qVJ4V1EOeQyQlJEhxYwmIZo11gyPnVBW
• Full logs reference: https://huggingface.co/datasets/cfahlgren1/Fable-5-traces