Prompt
I'm building a machine learning pipeline and need labeled data that is easy to version
Latest observation
Oct 16, 2026
OpenAI APIWeb search: off
A good fit is CSV/TSV with a small manifest and strict schema.
Why it’s easy to version
- Plain text, so diffs are readable in Git.
- Stable schema makes changes obvious.
- Works well with DVC, Git LFS, or simple Git if the dataset is small.
Recommended structure
data/train.csvdata/valid.csvdata/test.csvdata/labels.jsonorschema.yamlREADME.mddescribing label definitions and provenance
Best practices
- Use a unique
idcolumn for each example. - Keep labels in a separate, consistent column.
- Avoid changing row order; append or use stable sorting.
- Track metadata like source, annotator, and timestamp.
- If data is large, store files in object storage and version them with DVC.
Example CSV
id,text,label
1,"Great product",positive
2,"Very slow shipping",negative
3,"Okay overall",neutral
If you want more robust versioning
Consider:
- Parquet + DVC for larger tabular datasets
- JSONL for text/NLP tasks
- COCO/VOC-style JSON for computer vision annotations
If you tell me your data type—text, images, audio, or tabular—I can suggest the best labeling format and versioning workflow.