Hugging Face's logo

Datasets:
addyo07
/
vox-tier2-backup

Tasks:
Token Classification
Languages:
English
ArXiv:
Tags:
disfluency-detection
speech-refinement
vox
License:
Dataset card Files
xet
 Community
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Vox Tier-2 Dictation Refinement — Backup Snapshot (2026-10-04)

Failed-experiment parking snapshot. Work resumes next weekend with cleanup + parquet compression. Nothing in this repo is a release candidate. See docs/RECENT_WORK.md (ledger L3-HANDOFF-021) for the full handoff.

Single number that matters

pilot_v2_val_adjudicated.jsonl, roberta-base L0, threshold 0.9: 543/800 exact, 419/419 clean untouched, 124/381 disfluent fixed. Model deleted ~309 words, 22 wrong (need ≤6). Gate 2 (over_deletion < 1.0% AND harmful_span < 2.0%): over-deletion met (0.0%), harmful 7.12% — sole binding constraint.

Layout

Path What
data/corpus/ Frozen eval + training corpora (jsonl) + corpus_card.md (provenance/licences) + DATASET_SYNTHESIS_RECIPE.md
data/ladder/ L1/L2/L3 rungs (+ *_v2 reshaped) + *_meta.json + V_syn.jsonl held-out synthetic val
benchmarks/ One JSON per run (ft_*, ladder_*, rebase_*, threshold_extended, ladder_summary) + phaseB_ADJUDICATION.md
scripts/ Reusable harness: span_common.py, train_span.py, eval_span.py, run_ladder.py, assemble_ladder.py, adjudicate_val.py, qa_gate.py, compiler.py, schema.py, dedup.py, harvest_disfluency.py, corrupt_from_clean.py, …
docs/ GOAL.md (loop & gates), RECENT_WORK.md (sole ledger), AGENTS.md (invariants), spec/dictation-cleanup-engine.md (architectural spec)

Best-known-good training set

sandbox/corpus/pilot_v2_train.jsonl equivalent → data/corpus/pilot_v2_train.jsonl (3,199 rows: 1,423 real-human DISFLUENCY + 1,776 CLEAN). Evaluate ONLY on data/corpus/pilot_v2_val_adjudicated.jsonl (800 rows, CLEAN 419 / DISFLUENCY 381).

Reproduce the baseline 

venv/bin/python scripts/train_span.py --model FacebookAI/roberta-base \
  --out ckpt/roberta-L0 --train data/corpus/pilot_v2_train.jsonl \
  --seed 42 --epochs 8 --batch-size 16 --lr-encoder 3e-5 --lr-head 1e-3 \
  --weight-decay 0.01 --warmup-ratio 0.1 --max-len 128 --dtype float32
venv/bin/python scripts/eval_span.py --model ckpt/roberta-L0 --out benchmarks/repro.json
  • BIO labels O/B-DISFLUENCY/I-DISFLUENCY; label first sub-word only; is_split_into_words=True (NOT is_pretokenized — broken on transformers 5.x).
  • --dtype float32 mandatory (DeBERTa-v3 checkpoints are fp16; fp16 AdamW NaNs — see ledger L2-BAKEOFF-018).
  • Next-weekend plan (user-approved, NOT started): per-epoch val + keep-best in train_span.py FIRST, then rebuild AI data deterministically to disfluency_15k_v2-style shape (88/12 multi-span, no cue words, no em-dashes). Full detail: docs/RECENT_WORK.md §8 (L3-HANDOFF-021).

Licences / provenance

  • Real-disfluency harvest source: amaai-lab/DisfluencySpeech (Apache-2.0, arXiv:2406.08820 CC BY 4.0). Consumed fully (5,000 transcripts → 1,698 DISFLUENCY sentences). New real data needs a NEW source.
  • CLEAN pool sources + caps: see data/corpus/corpus_card.md.
  • Code in scripts/ is project harness (research snapshot, no separate licence file — treat as all-rights-reserved unless the owner adds one).

What is deliberately NOT in here

See EXCLUDED.md: 13 GB temp/ckpt ladder checkpoints, 2.6 GB downloads, production STT weights, venv, and anything downloadable from a link. This repo is the ~100 MB reproducible core.
Downloads last month
771
Total file size:
101 MB

Paper for addyo07/vox-tier2-backup