---
license: other
license_name: mindview-source-dependent-terms
license_link: https://huggingface.co/datasets/wangyanbo/MindView-Train/blob/main/DATA_LICENSE.md
language:
- en
tags:
- mindview
- multi-view
- spatial-reasoning
- image-generation
- supervised-fine-tuning
size_categories:
- 10K<n<100K
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: data/train.jsonl
---

# MindView-Train

**MindView-Train is the frozen supervised fine-tuning dataset used for MindView's BAGEL adaptation.** It contains **20,000 examples from 3,406 scenes** across ScanNet, ScanNet++, and ARKitScenes. Each example pairs six ordered reference RGB images and a spatial instruction with one held-out target RGB image. Source-aware scene identities are disjoint from [MindView-Bench](https://huggingface.co/datasets/wangyanbo/MindView-Bench).

[Code and training recipe](https://github.com/Yanbo-23/MindView) · [Benchmark](https://huggingface.co/datasets/wangyanbo/MindView-Bench) · [BAGEL-SFT weights](https://huggingface.co/wangyanbo/MindView-BAGEL-SFT) · [Collection](https://huggingface.co/collections/wangyanbo/mindview-6acf6b7a4a2e25efd51989f5)

## Contents

| Path | Purpose |
|---|---|
| `data/train.jsonl` | Original 20,000 training records |
| `data/schema.json` | Required fields and path conventions |
| `assets/rgb/*.tar` | 104,716 unique RGB images in 58 ordinary TAR archives |
| `manifests/` | Asset paths, archive membership, sizes, and checksums |
| `scripts/prepare_data.py` | Download verification and safe extraction |
| `RELEASE.json`, `SHA256SUMS` | Release inventory and file checksums |

Records preserve the `mindview_bagel_single_shot_v1` schema, identifiers, instructions, lineage, and input order. `shot=6`; `contexts["6"]` lists the reference paths, and `target_image` identifies the separate target. Repeated images are stored once. Paths are relative to the corresponding source dataset root.

## Download and prepare

Read [DATA_LICENSE.md](DATA_LICENSE.md) first. The first complete release is tagged `v1.0`; use this revision for both annotations and assets. Install the Hugging Face CLI if needed:

```bash
pip install -U huggingface_hub
hf download wangyanbo/MindView-Train \
  --repo-type dataset --revision "v1.0" \
  --local-dir ./MindView-Train
```

With Python 3.10 or later, verify the snapshot and extract into a separate directory:

```bash
python MindView-Train/scripts/prepare_data.py \
  --repo MindView-Train --verify-only
python MindView-Train/scripts/prepare_data.py \
  --repo MindView-Train --output ./mindview-data
```

Extraction restores `mindview-data/scannet/`, `scannetpp/`, and `arkitscenes/`. Set each runtime dataset root to its corresponding directory. Matching existing files are reused; conflicting files and unsafe archive entries are rejected. Packaging preserves original image bytes without applying training transforms.

For annotations only, use `load_dataset("wangyanbo/MindView-Train", "default", split="train", revision="v1.0")` from the `datasets` package. Records contain paths rather than automatically decoded images; download and extract the TAR assets separately. Pin the same revision for both.

## Training

The documented BAGEL recipe uses **10,000 optimizer steps on eight GPUs**. The runtime preserves context order and excludes target images from the inputs; it applies square cropping and BAGEL-specific resizing after decoding. Training additionally requires the matching BAGEL runtime, pretrained weights, model configuration, tokenizer, and VAE. See the project code for the complete recipe. Packaging checks do not constitute a new training run or GPU reproduction.

## Terms and contact

Source copyright and dataset-specific access and usage terms remain applicable. MindView's code license does not relicense source images; see [DATA_LICENSE.md](DATA_LICENSE.md).

Contact: Yihao Liu (`liuyihao@pjlab.org.cn`) and Jiwen Lu (`lujiwen@tsinghua.edu.cn`). Include the revision and record ID when reporting issues.