| name | dataset-curator |
| description | Curate, clean, and prepare datasets for AI training and evaluation. Invoke for: "clean this dataset", "prepare training data", "dataset curation", "deduplicate data", "label this data", "data quality", "prepare eval set", "filter bad examples".
|
| argument-hint | dataset path or description of data to curate |
| allowed-tools | Read, Write, Edit, Glob, Bash |
Skill: Dataset Curator — Training Data Preparation
Category: AI/ML Research
Role
Clean, deduplicate, and structure datasets for AI training, fine-tuning, or evaluation.
When to invoke
- Preparing training data for fine-tuning
- Building evaluation sets
- Cleaning scraped or collected data
- "this dataset is messy — clean it"
Instructions
- Load and profile the dataset: size, format, field distributions
- Remove duplicates (exact and near-duplicate using hashing or embeddings)
- Filter quality: remove empty, too-short, or clearly wrong examples
- Normalize: consistent format, encoding, whitespace
- Split: train/validation/test (80/10/10)
- Save cleaned version to
data/prompts/ or data/outputs/
- Document: data card with statistics and filtering decisions
Output format
## Dataset Curation Report — <dataset>
### Before: N examples, X% duplicates, Y% quality issues
### Filters Applied
### After: M examples (split: X train / Y val / Z test)
### Data Card
Example
/dataset-curator data/prompts/code_review_examples.jsonl — deduplicate and quality filter
