README.md

---
pretty_name: Medical-ROIs-K2.6
license: other
task_categories:
  - visual-question-answering
  - image-to-text
tags:
  - medical
  - radiology
  - pathology
  - roi
  - grounding
  - bounding-box
  - multimodal
  - sft
size_categories:
  - 10K<n<100K
language:
  - en
dataset_info:
  features:
    - name: id
      dtype: string
    - name: source
      dtype: string
    - name: teacher_model
      dtype: string
    - name: teacher_provenance
      dtype: string
    - name: images
      sequence: image
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
    - name: messages_mm
      dtype: string
    - name: problem
      dtype: string
    - name: solution
      dtype: string
    - name: prompt
      dtype: string
    - name: response
      dtype: string
    - name: seeing_regions
      list:
        - name: title
          dtype: string
        - name: bbox
          sequence: int64
    - name: seeing_json
      dtype: string
    - name: reasoning
      dtype: string
    - name: n_images
      dtype: int64
    - name: index
      dtype: int64
  splits:
    - name: train
      num_examples: 25134
    - name: validation
      num_examples: 2000
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
---

# Medical-ROIs-K2.6

Medical visual grounding SFT data: for each clinical VQA sample, a teacher model proposes **answer-supporting ROIs** (regions of interest) as 2D bounding boxes.

Teacher: **`moonshotai/Kimi-K2.6`**.  
Upstream images & QA: **`MBZUAI/medix-rl-data`**.

## How the data is generated

```
MBZUAI/medix-rl-data (train)
        │
        │  problem / solution / image / source / id
        ▼
  Teacher: moonshotai/Kimi-K2.6
  (vision + text; given question + gold answer)
        │
        │  ROI XML: <region><title>…</title><bbox>[x1,y1,x2,y2]</bbox></region>
        ▼
  JSONL (sft.jsonl)  →  stratified split
        │                 ├── sft_train.jsonl  (25,134)
        │                 └── sft_val.jsonl    (2,000)
        ▼
  prepare_hf: join JSONL text with medix-rl-data images
        │
        ▼
  Hugging Face DatasetDict  (this repo)
```

### 1. Source VQA pool

Rows come from [`MBZUAI/medix-rl-data`](https://huggingface.co/datasets/MBZUAI/medix-rl-data) (`train`, 51,335 examples) with four upstream sources:

| source | role |
|--------|------|
| `RadGenome/PMC-VQA` | PubMed Central figure VQA |
| `flaviagiammarino/path-vqa` | Pathology VQA |
| `BoKelvin/SLAKE` | Radiology VQA (SLAKE) |
| `flaviagiammarino/vqa-rad` | VQA-RAD |

Each row provides `id`, `image`, `problem` (question), `solution` (gold answer), and `source`.

### 2. Teacher ROI annotation

A multimodal teacher (`moonshotai/Kimi-K2.6`) is prompted with the image, clinical question, and gold answer, and asked to box the visual evidence that supports the answer.

- **Output format** (one or more blocks):

```text
<region><title>caption</title><bbox>[x1,y1,x2,y2]</bbox></region>
```

- **BBox convention:** integers in **[0, 1000]**, top-left origin; require `x1 < x2`, `y1 < y2`.
- Successful annotations are written to JSONL (`sft.jsonl`); failures are logged for repair/retry.
- This release keeps the successfully annotated subset (**27,134** rows, ≈53% of the upstream train pool).

### 3. Train / validation split

From `sft.jsonl`, a **source-stratified** random split (`seed=42`) produces:

| split | file | count |
|-------|------|------:|
| train | `sft_train.jsonl` | 25,134 |
| validation | `sft_val.jsonl` | 2,000 |

### 4. Hugging Face packaging

`medix_seeing_sft.prepare_hf` joins each JSONL row with the matching image(s) from `MBZUAI/medix-rl-data` by `id`, and materializes a `DatasetDict` with multimodal chat fields suitable for SFT loaders.

## Splits in this repo

| split | #examples |
|-------|----------:|
| `train` | 25,134 |
| `validation` | 2,000 |

### Source mix (approx.)

| source | train | validation |
|--------|------:|-----------:|
| RadGenome/PMC-VQA | 12,315 | 980 |
| flaviagiammarino/path-vqa | 9,535 | 759 |
| BoKelvin/SLAKE | 2,391 | 190 |
| flaviagiammarino/vqa-rad | 893 | 71 |

## Schema

| field | description |
|-------|-------------|
| `id` | Sample id (matches `medix-rl-data`) |
| `source` | Upstream VQA dataset name |
| `images` | Image list (`datasets.Image`) |
| `messages` | Chat turns with string content (`<image>` placeholders) |
| `messages_mm` | JSON string of multimodal content-list messages |
| `problem` / `solution` | Clinical question and gold answer |
| `prompt` / `response` | User prompt and assistant ROI XML |
| `seeing_regions` | Structured `{title, bbox}` list |
| `seeing_json` | JSON serialization of regions |
| `reasoning` | Teacher reasoning text (when available) |
| `teacher_model` | `moonshotai/Kimi-K2.6` |
| `teacher_provenance` | `distilled_from:moonshotai/Kimi-K2.6` |
| `n_images` | Number of images |
| `index` | Annotation-run index |

### Example assistant target

```text
<region><title>Right kidney</title><bbox>[130,380,470,720]</bbox></region>
<region><title>Left kidney</title><bbox>[570,420,840,750]</bbox></region>
```

## Load

```python
from datasets import load_dataset

ds = load_dataset("erow/Medical-ROIs-K2.6")
print(ds["train"][0]["problem"])
print(ds["train"][0]["response"])
ds["train"][0]["images"][0]  # PIL.Image
```

Or from a local `save_to_disk` folder:

```python
from datasets import load_from_disk
ds = load_from_disk("/path/to/hf_seeing_sft")
```

## Intended use

Supervised fine-tuning / distillation of medical VLMs for **visual grounding**: predicting ROIs that evidence the answer to a clinical question (complementary to answer-only VQA).

## Limitations

- ROIs are **teacher-generated**, not human-verified for every sample; some boxes may be loose, panel-misaligned, or weakly linked to the QA.
- Coverage is a subset of `medix-rl-data` (annotation failures / rate limits excluded).
- Not a clinical decision-support product; for research use only.

## Citation / provenance

- Upstream pool: [MBZUAI/medix-rl-data](https://huggingface.co/datasets/MBZUAI/medix-rl-data) and its constituent VQA datasets (PMC-VQA, PathVQA, SLAKE, VQA-RAD).
- Teacher: [Moonshot AI Kimi K2.6](https://huggingface.co/moonshotai/Kimi-K2.6) (via OpenAI-compatible inference).
- Pipeline: MediX seeing-SFT (`medix_seeing_sft` generate → split → `prepare_hf`).