README.md

---
license: mit
task_categories:
  - visual-question-answering
  - image-text-to-text
language:
  - en
pretty_name: Med Eval Data
size_categories:
  - 10K<n<100K
---

# Med Eval Data

This dataset contains evaluation data for the **Med** project. Its data format is the same as [`Med2026/Med_training_data`](https://huggingface.co/datasets/Med2026/Med_training_data), and it can be loaded with the same codebase from [`GAIR-NLP/Med`](https://github.com/GAIR-NLP/Med).

## Overview

Each example is stored in the same JSON / parquet schema as the training data, with the following top-level fields:

- `images`
- `data_source`
- `prompt`
- `ability`
- `reward_model`
- `extra_info`
- `agent_name`

This means the dataset is directly compatible with the data loading pipeline used in the Med codebase.

## Compatibility

This dataset has the **same format** as [`Med2026/Med_training_data`](https://huggingface.co/datasets/Med2026/Med_training_data).

You can use the same loading logic and preprocessing pipeline from:

- [`GAIR-NLP/Med`](https://github.com/GAIR-NLP/Med)

No format conversion is required.

## Data Split by File Naming

The evaluation data is divided into two settings according to the file name:

- Files with `single_turn_agent` in the filename correspond to **evaluation without tool use**
- Files with `tool_agent` in the filename correspond to **evaluation with tool use**

In other words:

- `*single_turn_agent*` → without-tool evaluation
- `*tool_agent*` → with-tool evaluation

## Data Format

Each sample is a JSON object with the following structure:

```python
{
  "images": [PIL.Image],
  "data_source": "vstar_bench_single_turn_agent",
  "prompt": [
    {
      "content": "<image>\nWhat is the material of the glove?\n(A) rubber\n(B) cotton\n(C) kevlar\n(D) leather\nAnswer with the option's letter from the given choices directly.",
      "role": "user"
    }
  ],
  "ability": "direct_attributes",
  "reward_model": {
    "answer": "A",
    "format_ratio": 0.0,
    "ground_truth": "\\boxed{A}",
    "length_ratio": 0.0,
    "style": "multiple_choice",
    "verifier": "mathverify",
    "verifier_parm": {
      "det_verifier_normalized": null,
      "det_reward_ratio": {
        "iou_max_label_first": null,
        "iou_max_iou_first": null,
        "iou_completeness": null,
        "map": null,
        "map50": null,
        "map75": null
      }
    }
  },
  "extra_info": {
    "answer": "A",
    "data_source": "vstar_bench_single_turn_agent",
    "id": "vstar_bench_0",
    "image_path": "direct_attributes/sa_4690.jpg",
    "question": "<image>\nWhat is the material of the glove?\n(A) rubber\n(B) cotton\n(C) kevlar\n(D) leather\nAnswer with the option's letter from the given choices directly.",
    "split": "test",
    "index": "0",
    "prompt_length": null,
    "tools_kwargs": {
      "crop_and_zoom": {
        "create_kwargs": {
          "raw_query": "What is the material of the glove?\n(A) rubber\n(B) cotton\n(C) kevlar\n(D) leather\nAnswer with the option's letter from the given choices directly.",
          "image": "PIL.Image"
        }
      }
    },
    "need_tools_kwargs": false
  },
  "agent_name": "single_turn_agent"
}