preference/README.md

---
pretty_name: IC-VCO Preference
task_categories:
- visual-question-answering
- image-text-to-text
language:
- en
---

# IC-VCO Preference

This is the `preference` configuration in the IC-VCO dataset package. It uses a
Parquet metadata file plus images stored in the shared package-level
`images/` directory. Image references are relative paths from this split's
metadata file to `../../images/<bucket>/...`.

## Schema

- `id`: stable sample id.
- `source_idx`: source row index.
- `prompt`: user prompt text.
- `chosen`: preferred response.
- `rejected`: dispreferred response.
- `images_file_names`: `[chosen_image, rejected_image]`, relative paths into the shared `../../images/<bucket>/` directories.
- `chosen_response_spans`: character spans in `chosen`.
- `rejected_response_spans`: character spans in `rejected`.

Response spans use zero-based character offsets with an exclusive end index.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("OPPOer/IC-VCO-Dataset", "preference")
```

For local repository training, `scripts/train.sh` defaults to the repo-root
`IC-VCO-Dataset` package. The preference stage loads this configuration
automatically.