CS230-lol/kev-jigsaw-toxicity-eval

Kev Jigsaw toxicity — raw evaluation outputs

Per-question predictions behind the results in github.com/CS230-lol/kev-tuning, plus the CONDA records Kev was scored on. Metrics, reports and the eval matrix live in the GitHub repo; this dataset holds the large files. Paths mirror the GitHub repo.

Models: CS230-lol/kev-0.8b-jigsaw-toxicity, CS230-lol/kev-4b-jigsaw-toxicity, and base jaredpalmer/kev-{0.8b,4b}. Training data: CS230-lol/JigsawToxicity (kev config).

Files

Path	What
runs/jigsaw-{0.8b,4b}/development/	Fine-tuned model on the 1,000 held-out Jigsaw records (7,000 questions)
runs/jigsaw-*/calibration/	Same on the 449 calibration records (temperature fit)
runs/jigsaw-*/baseline/{development,calibration}/	Base Kev on the same records
runs/jigsaw-*/regression/{finetuned,baseline}/	300 groups of Kev's public decision-v7 suite (forgetting check)
conda/results/<model>/<ctx|noctx>/	Zero-shot CONDA eval, 4 models × with/without context
conda/data/{ctx,noctx}.jsonl	CONDA CONDA_valid.csv (8,974 utterances) as Kev records: 18,380 records, 101,544 questions each
conda/data/index.jsonl	Record line → utterance index and question ids (same for both modes)
conda/data/gold.{label,seq.out}	CONDA gold intent and token tags, one utterance per line

Each eval directory has:

• predictions.jsonl: one line per record, {"id": "custom/<line>", "request_sha256", "prediction": {"probabilities": {question: {option: p}}}}, raw (temperature 1).
• rows.json: one entry per question, {id, question, keys, label, p, variant, ...} as written by kev.benchmark.evaluate_records.
• report.json: Kev's metric report.
• CONDA only: pred.label, pred.jigsaw.label, pred.seq.out (argmax predictions in CONDA's format) and result.json (UCA, U-F1, T-F1, JSA).

No Jigsaw text is stored here (record ids are line numbers in the kev config files). conda/data/ contains CONDA utterance text.

CONDA records

Built by conda/conda_to_kev.py from the CONDA repo. Per utterance:

• Record 1: a 4-way intent choice (explicit / implicit / action / other, from the paper's annotator guidelines) plus the 7 Jigsaw questions (their labels are placeholders: CONDA has no Jigsaw gold).
• Then one tok_<i> choice per token (toxic / character / dota / slang / pronoun / other), 8 per record; [SEPA] is never asked. Example words per class come from CONDA_train only.
• ctx: earlier utterances of the same conversation as THREAD: lines; noctx: the utterance alone. At run time, kev_conda_eval.py drops the oldest thread lines of 143 records (1,263 lines total) to fit Kev's 384-token state limit.

Licenses and credits

• Jigsaw / Civil Comments: CC0 1.0.
• CONDA (conda/data/): Weld et al., CONDA: a CONtextual Dual-Annotated dataset for in-game toxicity understanding and detection, Findings of ACL-IJCNLP 2021. The CONDA repo states no license; text is redistributed here for research reproducibility with attribution, and will be removed on request from the authors.
• Base models and scoring code: jaredpalmer/kev (Apache-2.0).

@inproceedings{weld-etal-2021-conda,
    title = "{CONDA}: a {CON}textual Dual-Annotated dataset for in-game toxicity understanding and detection",
    author = "Weld, Henry  and
      Huang, Guanghao  and
      Lee, Jean  and
      Zhang, Tongshu  and
      Wang, Kunze  and
      Guo, Xinghong  and
      Long, Siqu  and
      Poon, Josiah  and
      Han, Caren",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.213",
    doi = "10.18653/v1/2021.findings-acl.213",
    pages = "2406--2416",
}