Hugging Face's logo

Datasets:
rafmacalaba
/
datause-probe-v3

Modalities:
Tabular
Text
Formats:
json
Size:
10K - 100K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
datause-probe-v3 / README.md
rafmacalaba's picture
rafmacalaba
Upload folder using huggingface_hub
2811992 verified
preview code
|
Raw History Blame Contribute Delete
2.22 kB
metadata 
license: apache-2.0
configs:
  - config_name: splits
    data_files:
      - split: train
        path: train.jsonl
      - split: val
        path: val.jsonl
      - split: holdout
        path: holdout.jsonl
  - config_name: candidates
    data_files:
      - split: pool
        path: candidate_pool.jsonl
  - config_name: quarantine
    data_files:
      - split: excluded
        path: quarantine.jsonl

datause-probe-v3

Probe-v3 training data for the data-use mention probe: 29,346 labeled spans (26,618 campaign-v2 Luna single-judge + 2,728 gold-v2), doc-disjoint 70/15/15 splits, all six origins in every split, no passage crosses splits.

Files

  • train/val/holdout.jsonl — word-indexed spans (w_start/w_end over passage.split(" ")), label 1=DATA_MENTION 0=NON_MENTION, source camp2|gold, head_score = v2 probe score (triage feature, not label).
  • candidate_pool.jsonl — the 29,855 candidate spans the labels were drawn from: full rafmacalaba/datause-extracted-sample pool (2,729) + campaign-v2 draw (27,126), unified schema with passages. Slice-verified: 0 mismatches.
  • quarantine.jsonl — 8 spans excluded from training (7 unresolved echo + 1 bad_json), with reasons. Published, not silently dropped.
  • split_stats.json — per-split counts, docs, sources, origins.

Lineage

  1. Candidates from rafmacalaba/datause-extracted (gliner2), scored by rafmacalaba/gliner-datause-v2-probe.
  2. Campaign v2: origin x specificity x 0.1-decile draw (144 cells, floor 50 + sqrt-proportional, seed 0) → 27,126 single-span keyed Luna requests (v2.4 rules + v2.5 presentation, prompt v2.5-draft-2026-09-06), one run, no consensus. 500 gold-v2 controls interleaved (94.0% judge-vs-gold agreement).
  3. Echo re-attribution: 965/972 span-echo mismatches resolved as whitespace/punctuation normalization (single-span requests, reasons audited); 7 quarantined.
  4. TrainingDataset.validate(): 29,346 valid / 0 invalid (strict verbatim).
  5. build_probe_v3_data.py: doc-disjoint splits; controls enter once via gold.

Labels are a Luna distillation target, not human gold. The 473 human-adjudicated spans (annotator190/jdc283) are NOT in this dataset — they remain the external check.