collusion-watch / data /README.md
munip's picture
Show Collusion-Watch article, runs ZIP, and project files
d765911 verified
|
Raw History Blame Contribute Delete
2.52 kB
---
pretty_name: Collusion-Watch Aya Market Dataset
language:
- en
task_categories:
- text-classification
- text-generation
tags:
- multi-agent-systems
- collusion-monitoring
- synthetic-data
size_categories:
- n<1K
---
# Collusion-Watch Aya Market Dataset
Reviewed release of 112 simulated double-auction episodes, their evidence
logs, the seed-split M2 transcript dataset, the train-only CuratorKIT export,
and monitor sweep results. See the [project Space](https://huggingface.co/spaces/munip/collusion-watch)
for the complete write-up, source snapshot, and model links.
This repository is included in the public
[Collusion-Watch dataset collection](https://huggingface.co/collections/munip/collusion-watch-datasets-6afae07e8895a2910ce66914).
## Contents
- `aya_market_base_a2prompt/`: 112 selected episodes, `episodes.csv`,
configuration, artifact manifest, and run metadata. Each episode directory
includes `episode.json` and `evidence.jsonl`.
- `aya_market_base_a2prompt_runs.zip`: ZIP of that same filtered 112-episode
package.
- `aya_m2_dataset_41seed30_v1/`: 112 M2 transcript examples and split
manifest (78 train, 18 validation, 16 test; grouped by seed).
- `aya_m2_curated_41seed30_v1/`: 78-example train-only CuratorKIT export and
provenance artifacts; zero rows were rejected.
- `aya_m2_monitor_41seed30_comparison_v1/`: original sweep outputs.
- `aya_m2_monitor_41seed30_comparison_v1_reconciled/`: offline
re-aggregation of the same saved decisions, including three unparseable
base-judge responses per judge-using monitor arm.
The ZIP is intentionally not an archive of the entire local `runs/` folder.
It excludes unrelated experiments, private monitor keys/anchors, credentials,
and unreviewed raw captures.
## Scope and limitations
The data is synthetic and English-only. It does not demonstrate real-world
collusion, multilingual safety, or safety-preserving fine-tuning. The held-out
test set has only four benign episodes; A5 is present in training and test.
The benchmark is for research and reproducibility, not operational decisions.
## License
No separate reuse license is assigned to this dataset in this release. Public
hosting does not itself grant permission to reuse it. The upstream Tiny Aya
Fire and Tiny Aya Global model cards list CC-BY-NC-4.0 and gated access; review
those model terms and obtain any required access before use.
## Citation
Please cite the project article and repository. A formal citation and DOI may
be added when available.