---
license: apache-2.0
task_categories:
- image-to-text
- visual-question-answering
language:
- en
---

# Prime-X12 30B training data

This repository contains the verified Prime-X12 multimodal training package used
for `clonelabs/prime-x12-30b`.

## One-command download

```bash
huggingface-cli download clonelabs/prime-x12-30b-data \
  archives/prime-x12-30b-data-aa5447e6.zip \
  --repo-type dataset \
  --local-dir .
unzip archives/prime-x12-30b-data-aa5447e6.zip -d prime-x12-30b-data
```

Verify the download before extracting:

```bash
huggingface-cli download clonelabs/prime-x12-30b-data \
  archives/prime-x12-30b-data-aa5447e6.zip.sha256 \
  --repo-type dataset \
  --local-dir .
(cd archives && sha256sum -c prime-x12-30b-data-aa5447e6.zip.sha256)
```

The ZIP is stored without recompressing images (`zip -0`) and contains:

- `data/train.jsonl` — 24,929 examples
- `data/validation.jsonl` — 100 examples
- `images/` — every image referenced by the two splits
- `image_manifest.jsonl` — path, byte size, and SHA-256 for each image
- `metadata.json`
- `source_split_manifest.json`

Desktop-1, Desktop-2, and OpenCUA sources are mixed with a fixed seed. Targets
use the standard `<think>` tag.