Datasets:
dataset card: composition, provenance, AGPL-3.0 inherited from upstream
Browse files
README.md
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: agpl-3.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
task_categories:
|
| 6 |
+
- text-generation
|
| 7 |
+
tags:
|
| 8 |
+
- agent-traces
|
| 9 |
+
- synthetic
|
| 10 |
+
- coding
|
| 11 |
+
- reasoning
|
| 12 |
+
size_categories:
|
| 13 |
+
- 10K<n<100K
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# Parable corpus v2
|
| 17 |
+
|
| 18 |
+
The training corpus behind the [Parable](https://huggingface.co/collections/AnkitAI/parable)
|
| 19 |
+
model series: genuine multi-step agent-session traces, deduplicated,
|
| 20 |
+
decontaminated and mixed with general-instruction replay data.
|
| 21 |
+
|
| 22 |
+
Released so that the models trained on it can be checked rather than taken
|
| 23 |
+
on trust. Method and measurements:
|
| 24 |
+
[doi:10.5281/zenodo.21676407](https://doi.org/10.5281/zenodo.21676407).
|
| 25 |
+
|
| 26 |
+
## Composition
|
| 27 |
+
|
| 28 |
+
| Source | License | Role |
|
| 29 |
+
|---|---|---|
|
| 30 |
+
| [Glint-Research/Fable-5-traces](https://huggingface.co/datasets/Glint-Research/Fable-5-traces) | AGPL-3.0 | Claude Fable 5 agent sessions |
|
| 31 |
+
| [Roman1111111/gpt5.5-terminal](https://huggingface.co/datasets/Roman1111111/gpt5.5-terminal) | MIT | terminal-session transcripts |
|
| 32 |
+
| OpenCoder educational_instruct, tulu-3 mix | see upstream | general-instruction replay (~30%) |
|
| 33 |
+
|
| 34 |
+
Processing applied: session reconstruction, secret scrubbing, exact and
|
| 35 |
+
fuzzy deduplication, and fuzzy train/test decontamination against the
|
| 36 |
+
benchmarks reported in the paper, with post-hoc containment measured rather
|
| 37 |
+
than assumed.
|
| 38 |
+
|
| 39 |
+
## Licensing
|
| 40 |
+
|
| 41 |
+
**This corpus is released under AGPL-3.0**, inherited from the
|
| 42 |
+
Fable-5-traces portion, which is the most restrictive upstream term. Reusing
|
| 43 |
+
this corpus carries that obligation forward.
|
| 44 |
+
|
| 45 |
+
Because the traces originate from third-party assistants, those providers'
|
| 46 |
+
terms may also apply to downstream training and distillation independently
|
| 47 |
+
of this license. Confirm your use aligns with them before building on this
|
| 48 |
+
commercially.
|
| 49 |
+
|
| 50 |
+
## Citation
|
| 51 |
+
|
| 52 |
+
```bibtex
|
| 53 |
+
@misc{aglawe2026agenttrace,
|
| 54 |
+
author = {Aglawe, Ankit},
|
| 55 |
+
title = {Agent-Trace Fine-Tuning of Small Language Models under Constrained Compute},
|
| 56 |
+
year = {2026},
|
| 57 |
+
publisher = {Zenodo},
|
| 58 |
+
doi = {10.5281/zenodo.21676407},
|
| 59 |
+
url = {https://doi.org/10.5281/zenodo.21676407}
|
| 60 |
+
}
|
| 61 |
+
```
|