Ximeng0831/NaviDrive-Reasoning

---
license: apache-2.0
task_categories:
- image-text-to-text
- text-generation
language:
- en
tags:
- nuscenes
- driving-reasoning
configs:
- config_name: gemini
data_files:
- split: train
path: data/gemini_train.jsonl
- split: validation
path: data/gemini_val.jsonl
- config_name: qwen_8b
data_files:
- split: train
path: data/qwen_8b_train.jsonl
- split: validation
path: data/qwen_8b_val.jsonl
- config_name: qwen_32b
data_files:
- split: train
path: data/qwen_32b_train.jsonl
- split: validation
path: data/qwen_32b_val.jsonl
- config_name: qwen_8b_mini
data_files:
- split: train
path: data/mini_qwen_8b_train.jsonl
- split: validation
path: data/mini_qwen_8b_val.jsonl
---

# NaviDrive-Reasoning

[](https://arxiv.org/abs/2603.07901)
[](https://huggingface.co/Ximeng0831/NaviDrive-Qwen3-VL-2B-SFT)
[](https://huggingface.co/datasets/Ximeng0831/NaviDrive-Reasoning)
[](https://github.com/TAMU-CVRL/NaviDrive)

This dataset contains reasoning, perception, and action explanations generated for the NuScenes dataset using multiple LLMs.

## Dataset Summary
- **Gemini**: Reasoning generated by the [Gemini-2.5-Flash model]((https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash)).
- **Qwen-8B**: Reasoning generated by the [Qwen-8B model]((https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)).
- **Qwen-32B**: Reasoning generated by the [Qwen-32B model]((https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct)).
- **Mini**: A subset specifically for the nuScenes-mini dataset.

> **Qwen-32B** is the primary dataset used for model training.

## Data Structure
Each `.jsonl` file follows a consistent schema. Key fields include:

| Field | Description |
| :--- | :--- |
| `token` | Unique identifier for the nuScenes sample. |
| `command` | High-level driving intent (e.g., `<Keep_Straight>`, `<Turn_Right>`). |
| `wp_future` | A sequence of (x, y, \theta) future waypoints for the ego-vehicle. |
| `reasons` | A list containing the **Perception**, **Action**, and **Reasoning** breakdown. |
| `image_paths` | Relative paths to the corresponding nuScenes camera sensors. |
> **Note**
> Images are not included in the dataset—only relative image paths are provided.
> To use image inputs, please download the [nuScenes Dataset](https://www.nuscenes.org/nuscenes#download) and specify the dataset root path to properly load the images.

## How to Load
```python
from datasets import load_dataset

# Example: Load the NaviDrive-Reasoning Dataset
train_dataset = load_dataset("Ximeng0831/NaviDrive-Reasoning", "qwen_32b", split="train")
val_dataset = load_dataset("Ximeng0831/NaviDrive-Reasoning", "qwen_32b", split="validation")

# Access the first sample
print(train_dataset[0]["reasons"])
```

## Acknowledgements
This dataset is based on the [nuScenes dataset](https://www.nuscenes.org/).