ErfanRou/callcc-new-data-labelling

Call-centre labelling — current V7 anonymous five-output assignments

This repository uses manual access approval. Customer audio and transcripts stay within the authorized project. Give each annotator only their V7 ZIP through the team's private channel; the ZIP works offline.

Download your assignment

Person	Assignment
Arash	Download ZIP
Faraz	Download ZIP
Hossein	Download ZIP
Mahshid	Download ZIP
Mehrzad	Download ZIP
Mohammad Sadegh	Download ZIP
Mohammadreza	Download ZIP
Pedram	Download ZIP
Zahra	Download ZIP

Unzip the file, read START_HERE.txt, then open page_01.html. Pages contain about 12 minutes of audio; difficult clips take longer than their audio duration.

Anonymous five-output workflow

1. Listen to the foreground speaker before choosing anything.
2. Every clip shows five unnamed outputs. Their order is randomized independently per clip; output numbers do not identify a model.
3. All five have the same controls: mark the full output correct, start editing from it, or click one highlighted part to place only that part in the shared editor. Typing elsewhere is preserved. A loaded draft alone remains pending.
4. Preserve colloquial words, fillers, repeats and spoken number groups. For «چهل و هشت، نود و سه», write 48 93; preserve leading zeros and individually spoken digits. Never infer a number from context.
5. Use overlap, unusable, non-Persian and notes when appropriate.
6. Export JSONL after every session, keep dated copies, and send the latest export privately to the project owner. Export files contain anonymous out1–out5 references; the private merge key restores provenance afterward.

Keyboard: Esc play/pause, Shift+Esc back three seconds, 1–5 marks an output correct, E focuses the editor, and Ctrl/Cmd+Enter leaves it. V7 has a new page/storage namespace and schema. Preserve every earlier export.

Why these clips were selected

The final set contains 4,370 clips. Every selected clip has all five usable drafts, a meaningful production-versus-Transcribe contrast, and at least three meaningful production-versus-alternative contrasts. Numeric disagreement or at least two normalized word edits with distance >= 0.20 counts as meaningful; failed calls never count as votes.

The selection includes both numeric and lexical disagreements and balances lower-SNR, clearer, and unknown-SNR clips. Lower values generally mean noisier audio; this is an MP3/VAD proxy rather than calibrated clean-reference SNR.

Disagreement guides sampling; humans establish the label. Gemini-family models are correlated, so agreement is not treated as an independent majority vote.

Project-owner model-output archive

The eligible pool contains 24,398 clips / 109.162 audio hours. All available paid machine drafts and explicit statuses remain preserved under V5. These are machine outputs, not human labels.

Model	Attempted audio hours	Usable draft hours
prod	109.162	109.162
soniox	109.162	109.162
google/gemini-3.5-transcribe	109.162	108.003
google/gemini-3.8-flash	55.508	47.232
google/gemini-4-argon	49.795	44.228

• transcripts.parquet and transcripts.jsonl: saved model-output archive.
• audio/: MP3 annotation previews joined by id.
• V7 verification and selection summary.
• previous V6 guide.

The private output-to-model mapping is not in this repository. Retrieve original lossless source audio after labels return and before training. Keep the reserved evaluation set out of training.