Datasets:
license: agpl-3.0
language:
- en
task_categories:
- text-generation
tags:
- agent-traces
- synthetic
- coding
- reasoning
size_categories:
- 10K<n<100K
Parable corpus v2
The training corpus behind the Parable model series: genuine multi-step agent-session traces, deduplicated, decontaminated and mixed with general-instruction replay data.
Released so that the models trained on it can be checked rather than taken on trust. Method and measurements: doi:10.5281/zenodo.21676407.
Composition
| Source | License | Role |
|---|---|---|
| Glint-Research/Fable-5-traces | AGPL-3.0 | Claude Fable 5 agent sessions |
| Roman1111111/gpt5.5-terminal | MIT | terminal-session transcripts |
| OpenCoder educational_instruct, tulu-3 mix | see upstream | general-instruction replay (~30%) |
Processing applied: session reconstruction, secret scrubbing, exact and fuzzy deduplication, and fuzzy train/test decontamination against the benchmarks reported in the paper, with post-hoc containment measured rather than assumed.
Licensing
This corpus is released under AGPL-3.0, inherited from the Fable-5-traces portion, which is the most restrictive upstream term. Reusing this corpus carries that obligation forward.
Because the traces originate from third-party assistants, those providers' terms may also apply to downstream training and distillation independently of this license. Confirm your use aligns with them before building on this commercially.
Citation
@misc{aglawe2026agenttrace,
author = {Aglawe, Ankit},
title = {Agent-Trace Fine-Tuning of Small Language Models under Constrained Compute},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.21676407},
url = {https://doi.org/10.5281/zenodo.21676407}
}