luca0621/amex-gelab-448

---
pretty_name: AMEX SFT
task_categories:
- image-text-to-text
- text-generation
tags:
- gui
- mobile
- navigation
- multimodal
- ge-lab
size_categories:
- 10K<n<100K
---

# AMEX SFT

This dataset is a packaged export of the local `amex_sft` directory for uploading to the Hugging Face Hub as a dataset repository.

## Source

- Source dataset roots:
- `/home1/irteam/data-vol1/amex_sft_hf_448` (3046 trajectories)
- Number of trajectory folders: `3046`
- Number of tar shards: `61`
- Trajectories per shard: `50`

## Layout

- `shards/*.tar`: tar shards containing trajectory folders
- `manifest.jsonl`: trajectory-to-shard index
- `dataset_info.json`: high-level metadata

Each tar shard preserves the original trajectory folder layout. For example:

```text
<source_root>/<trajectory_id>/
ui_structure.json
ui_structure_layer.json
trajectory_assets_manifest.json
action_coord/...
extracted_assets/...
```

## Why tar shards?

The raw source contains a very large number of small PNG files. Packaging them into tar shards makes Hub upload and downstream download much more reliable for large-scale storage.

## Notes

- This repository is intended for storage and reuse of the packaged dataset.
- If you want Hub-native previews and lighter access patterns, consider a future Parquet/WebDataset conversion.