README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- visual-question-answering
- video-text-to-text
language:
- en
tags:
- medical
- chest-ct
- radiology
- ct-rate
- dpo
size_categories:
- 10K<n<100K
pretty_name: CT-RATE-AB
configs:
- config_name: AB
  data_files:
  - split: train
    path: viewer/train.parquet
  - split: valid
    path: viewer/valid.parquet
- config_name: AB_DPO
  data_files:
  - split: train
    path: viewer/dpo.parquet
---

# CT-RATE-AB

Vision-language annotations for chest CT abnormality reporting, derived from
[CT-RATE](https://huggingface.co/datasets/ibrahimhamamci/CT-RATE) and
formatted in the LLaVA conversation schema. Includes both an SFT split
(train / valid) and a DPO preference set.

> ⚠️ Research use only. Not a medical device. Do not use for clinical decisions.

## Splits

| Subset | # Samples | Purpose                  |
|--------|-----------|--------------------------|
| train  | 46,709    | SFT training             |
| valid  | 3,039     | Validation               |
| DPO    | 46,709    | DPO preference fine-tuning |

The DPO entries share `id` values with the SFT train set (the `chosen`
response equals the SFT `gpt` value); train and valid splits inherit the
official CT-RATE partition (no patient overlap).

## ⚠️ Two file formats in this repo

This repository ships **two parallel views** of the same data:

| Purpose                | Files                                    | Format |
|------------------------|------------------------------------------|--------|
| **Training (use these)** | `CT-RATE-AB-train.json`, `CT-RATE-AB-valid.json`, `CT-RATE-AB-DPO.json` | Raw LLaVA-format JSON |
| **Dataset Viewer only**  | `viewer/train.parquet`, `viewer/valid.parquet`, `viewer/dpo.parquet`     | Simplified parquet    |

**For training, always load the JSON files.** They preserve the full LLaVA
conversation schema (`videos`, `conversations` with `<video>` token, etc.)
expected by standard multimodal training pipelines.

The parquet files in `viewer/` exist **only so the Hugging Face Dataset
Viewer shows the report text cleanly** — they strip out the `videos`
filename column (no videos are hosted here) and the boilerplate human
prompt, leaving just `id` and the report. **Do not use parquet for
training.** The two views are guaranteed identical in content; they
differ only in field layout.

## Data format (the JSON files)

Standard LLaVA video-conversation schema.

**`CT-RATE-AB-train.json` / `CT-RATE-AB-valid.json`** (SFT):
```json
{
  "id": "train_1_a_1",
  "videos": ["train_1_a_1.mp4"],
  "conversations": [
    {"from": "human", "value": "<video>\nGenerate a report for this patient."},
    {"from": "gpt",   "value": "Lung: ...\n\nMediastinum: ...\n\n..."}
  ]
}
```

**`CT-RATE-AB-DPO.json`** (DPO preference pairs):
```json
{
  "id": "train_5251_a_1",
  "videos": ["train_5251_a_1.mp4"],
  "conversations": [
    {"from": "human", "value": "<video>\nGenerate a report for this patient."}
  ],
  "chosen":   {"from": "gpt", "value": "Lung: ... atelectatic changes ..."},
  "rejected": {"from": "gpt", "value": "Lung: ... ground-glass opacities ..."}
}
```

The assistant response is a structured report organized by anatomy:
*Lung*, *Trachea and Bronchie*, *Mediastinum*, *Bone*, *Abdomen*, *Others*.
The human prompt is identical across all samples:
`<video>\nGenerate a report for this patient.`

## Obtaining the CT volumes

This repository contains **only annotations**. To obtain the corresponding CT
data, request access to CT-RATE at
<https://huggingface.co/datasets/ibrahimhamamci/CT-RATE>. The `id` field in
each entry maps directly to the CT-RATE volume identifier (and to the mp4
filename listed in `videos`, if you render videos locally).

## License

Released under **CC BY-NC-SA 4.0**, consistent with the upstream CT-RATE
license.