DanielDDDS/recipe-modifications

---
license: mit
---

# Hebrew Recipe Modification Dataset

## Overview

10,058 Hebrew comment threads from YouTube cooking channels, annotated for recipe modification extraction using a three-pass Teacher-Student distillation approach.

## Task

Token-level BIO tagging to extract recipe modifications from Hebrew user comments. Four modification aspects: SUBSTITUTION, QUANTITY, TECHNIQUE, ADDITION.

## Dataset Structure

### Raw Data
- **threads.jsonl** — 10,058 comment threads (top comment + replies) from 17 Hebrew cooking channels
- **teacher_output.jsonl** — Silver labels from three-pass annotation with majority vote
- **final_stats.json** — Labeling statistics

### Processed (BIO-tagged, ready for training)
- **processed/train.jsonl** — 8,129 examples (1,066 positive)
- **processed/val.jsonl** — 1,016 examples (134 positive)
- **processed/test.jsonl** — 1,012 examples (128 positive)
- **processed/label2id.json** — 9 BIO labels
- **processed/id2label.json** — Reverse mapping
- **processed/stats.json** — Preprocessing statistics

## Labeling Pipeline

| Pass | Model | Temperature | Role |
|------|-------|-------------|------|
| 1 | Gemini 3.1 Flash Lite | 0.1 | Primary annotator |
| 2 | Gemini 3.1 Flash Lite | 0.3 | Intra-annotator consistency |
| 3 | Qwen 3 235B (Cerebras) | 0.1 | Inter-annotator validation |

### Agreement

| Vote Method | Count | % |
|-------------|-------|---|
| Unanimous (3/3) | 8,907 | 88.6% |
| Majority (2/3) | 1,142 | 11.4% |
| Manual review | 9 | 0.1% |

### Final Labels

| Status | Count | % |
|--------|-------|---|
| With modification | 1,230 | 12.2% |
| No modification | 8,828 | 87.8% |

## Label Schema (BIO, 9 labels)
``````
O, B-SUBSTITUTION, I-SUBSTITUTION, B-QUANTITY, I-QUANTITY, B-TECHNIQUE, I-TECHNIQUE, B-ADDITION, I-ADDITION
``````

## Intended Use

Training Hebrew token classification models (e.g., AlephBERT) for recipe modification extraction via knowledge distillation.

## Citation

If you use this dataset, please cite our project report.
"@ | Out-File -Encoding utf8 data\hf_readme.md

huggingface-cli upload DanielDDDS/recipe-modifications data/hf_readme.md README.md --repo-type dataset