lorde-data / README.md
Oxlino's picture
Dataset card: list only files that exist; status, licences and caveats
a4e4b80 verified
|
Raw History Blame Contribute Delete
3.9 kB
metadata
license: other
license_name: cc-by-nc-4.0-and-cohere-aup-and-cc-by-sa-4.0-mixed
license_link: https://cohere.com/c4ai-cc-by-nc-license
language:
  - en
  - hi
  - fr
  - ar
  - es
  - pt
  - id
  - tr
tags:
  - llm-judge
  - data-curation
  - curatorkit
  - lexsi-hackathon
  - non-commercial
pretty_name: Lorde judge data

Lorde judge data

Licence: mixed, non-commercial. Teacher labels and generated QA come from Aya Expanse (CC-BY-NC-4.0 + Cohere Labs Acceptable Use Policy). This dataset also includes CC BY-SA 4.0 content: Wikipedia excerpts (source passages, and text derived from them); the CC BY-SA 4.0 terms (attribution, share-alike) continue to apply to that content. Details below. Un-gated for the Lexsi hackathon; using the data means accepting these terms.

The data, labels, CuratorKIT plug-in runs and result files behind Oxlino/lorde-tiny-aya. This repository holds exactly these files:

Path Contents
pool.jsonl 6720 judge items: uid, kind (reward/grounding), lang, group, variant, item fields
labels.jsonl teacher (CohereLabs/aya-expanse-8b) replies to CuratorKIT's exact prompts, parsed by CuratorKIT
sft/train.jsonl, sft/val.jsonl 5296 / 309 SFT rows (canonical JSON targets)
test.jsonl 937 held-out items with teacher labels (grouped by prompt / passage)
human_sample.jsonl 200 blind test items (opaque ids, shuffled) for hand labelling with the gate rubric; no hand labels have been collected yet, so none are published
downstream_raw.jsonl, downstream_eval.jsonl downstream QA set and its held-out eval set
gated/<judge>/ CuratorKIT outputs per judge (base, lorde, lorde+so, teacher): passed/rejected; manifest and lexsi_provenance.json for base, lorde, lorde+so, teacher
gated/ungated/passed.jsonl the whole downstream raw set, as the ungated baseline of the downstream test
bench/ Lorde Bench sets: adversarial-100, golden-300
results/ result files (JSON and Markdown): downstream, judge_eval, labels_summary, plugin_eval; per-judge predictions in results/preds/
pool_stats.json, splits_summary.json pool and split statistics

Status and caveats

  • Labels: CohereLabs/aya-expanse-8b (self-hosted) labels every item; the planned c4ai-aya-expanse-32b calibration slice was not run; no hand labels have been collected yet, so human agreement is not measured.
  • The teacher is the reference for every agreement number, and it is lenient on grounding (see the teacher audit in the model card): agreement with it measures imitation, not correctness.
  • Near-duplicates: 17 of 348 test reward prompts have a ≥ 90%-similar training prompt (normalized text), and 12 test gold responses equal a training one.
  • If you are hand-labelling human_sample.jsonl, do not consult this dataset while labelling: pool.jsonl and test.jsonl hold the same texts with their variant, so looking items up would de-blind the labels.

Sources and licences

  • Prompts and human answers: CohereLabs/aya_dataset (Apache-2.0); its language labels are used as given.
  • Passages: Wikipedia via wikimedia/wikipedia (CC BY-SA 4.0; older revisions CC BY-SA 3.0, most text also GFDL). The excerpts in this dataset, and text derived from them, stay under CC BY-SA: attribute Wikipedia's contributors and share adaptations under the same licence.
  • Teacher labels, generated QA and translations: Aya Expanse 8B; student and closed-book answers: Tiny Aya Global (both CC-BY-NC-4.0 + Cohere Labs Acceptable Use Policy): non-commercial research use.
  • Produced with CuratorKIT and AuditKIT (Lexsi Labs LSAL, non-commercial source-available). Third-party terms: THIRD_PARTY_LICENSES.md in the GitHub repo.