---
language: en
license: apache-2.0
task_categories:
- visual-question-answering
- object-detection
tags:
- uav
- drone
- reasoning
- vqa
- chain-of-thought
- qwen-vl
- counter-uas
configs:
- config_name: v2
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# UAV Drone Reasoning VQA Dataset v2

Multimodal VQA dataset for counter-UAS training with rich chain-of-thought reasoning.

## v2 Improvements
- Negative and hard-negative samples (birds/aircraft distractors)
- Category-labeled QA pairs
- Richer multi-sentence thinking traces
- Qwen-VL chat `messages` format with `` blocks
- Stratified train/validation/test splits
- Duplicate-question downweighting via `dup_weight`

## Stats
- Records: 14,909
- QA pairs: 98,837
- Unique questions: 14,590
- Sources: {'kaggle_v1': 9909, 'negative': 2500, 'hard_negative': 2500}

## Fields
| Field | Description |
|-------|-------------|
| `image` | PIL-ready surveillance image |
| `messages` | Qwen chat JSON (user question + assistant thinking/answer) |
| `ground_truth` | Pixel xyxy boxes + class |
| `ground_truth_norm` | 0-1000 normalized boxes |
| `categories` | QA category labels |
| `dup_weight` | Sampling weight for deduplicated training |
| `source` | kaggle_v1 / negative / hard_negative |

## Usage
```python
from datasets import load_dataset
import json

ds = load_dataset("yunus-emre/uav-drone-reasoning-vqa", "v2")
row = ds["train"][0]
messages = json.loads(row["messages"])
```