---
language:
- en
tags:
- multimodal
- instruction-tuning
- genllava
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: data/*/*.parquet
- config_name: llava_text
  data_files:
  - split: train
    path: data/llava_text/*.parquet
- config_name: llava_understanding
  data_files:
  - split: train
    path: data/llava_understanding/*.parquet
- config_name: lvis
  data_files:
  - split: train
    path: data/lvis/*.parquet
- config_name: editing
  data_files:
  - split: train
    path: data/editing/*.parquet
- config_name: generation
  data_files:
  - split: train
    path: data/generation/*.parquet
- config_name: lrv
  data_files:
  - split: train
    path: data/lrv/*.parquet
- config_name: laion_gpt4v
  data_files:
  - split: train
    path: data/laion_gpt4v/*.parquet
- config_name: sharegpt4v
  data_files:
  - split: train
    path: data/sharegpt4v/*.parquet
- config_name: docvqa
  data_files:
  - split: train
    path: data/docvqa/*.parquet
- config_name: synthdog
  data_files:
  - split: train
    path: data/synthdog/*.parquet
- config_name: chartqa
  data_files:
  - split: train
    path: data/chartqa/*.parquet
- config_name: dvqa
  data_files:
  - split: train
    path: data/dvqa/*.parquet
- config_name: ai2d
  data_files:
  - split: train
    path: data/ai2d/*.parquet
---
# Generative Visual Instruction Tuning

Release of dataset for **Generative Visual Instruction Tuning**. 

## Contents

| Task | Rows |
|---|---:|
| text | 40,688 |
| understanding | 1,221,051 |
| editing | 200,000 |
| generation | 558,120 |

| Component | Packaged rows |
|---|---:|
| llava_text | 40,688 |
| llava_understanding | 624,207 |
| lvis | 213,690 |
| editing | 200,000 |
| generation | 558,120 |
| lrv | 79,798 |
| laion_gpt4v | 11,576 |
| sharegpt4v | 102,025 |
| docvqa | 39,461 |
| synthdog | 65,000 |
| chartqa | 22,881 |
| dvqa | 50,000 |
| ai2d | 12,413 |

Only training data are packaged. 

## Loading

Load the complete mixture or one component:

```python
from datasets import load_dataset
data = load_dataset("jefehern/genllava", "default", split="train", streaming=True)
editing = load_dataset("jefehern/genllava", "editing", split="train", streaming=True)
example = next(iter(editing))
source_image = example["input_images"][0]
edited_image = example["target_image"]
```
## Record schema and training use
`task` is `text`, `understanding`, `generation`, or `editing`. `messages` contains role/content pairs. Understanding rows provide `input_images`; generation rows provide only `target_image`; editing rows provide one source image and a separate target. Text rows have neither.

`<image>` identifies a supplied input image; `<image_output>` marks image supervision. These are semantic dataset markers.  This dataset preserves full released conversations and exact OCR labels; it does not apply the paper's 2,048-token text truncation or reproduce an unspecified task-token placement.

## Recipe decisions and exclusions
- The 40,688 text-only ShareGPT conversations were partitioned from LLaVA's 665,298 records and counted once.
- All 558,128 released generation pairs and all 102,025 ShareGPT4V caption rows were considered. Duplicate or invalid records were removed and counts are reported above.
- Editing selects 200K source rows, SynthDoG 65K, ChartQA 23K, and DVQA 50K by seed-7 BLAKE2b ranking. LRV selects 80K unique source QA records from the two upstream coordinate-free releases plus chart annotations. No rows are added back after quality exclusions or exact deduplication.
- DocVQA uses its 39,463-question training split. AI2D uses 12,413 records whose 3,246 image IDs have zero overlap with the official test split's 814 image IDs. Validation/test questions are not added to reach the paper's rounded targets.
- LAION has 12,356 released metadata rows. **11,576** images were recovered; **780** remain unavailable after mirror recovery and bounded URL retries.
- Deduplication uses exact task/messages/input-image hashes/target-image hash. It retains different questions about the same image. Global duplicates and held-out image matches removed during packaging are listed in `provenance/package_exclusions.jsonl`.

## Sources
- [LLaVA generation / pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain)
- [LLaVA instruction mixture, including ShareGPT](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K)
- [LVIS-Instruct4V](https://huggingface.co/datasets/X2FD/LVIS-Instruct4V)
- [InstructPix2Pix](https://huggingface.co/datasets/timbrooks/instructpix2pix-clip-filtered)
- [LRV coordinate-free annotations and chart data](https://github.com/FuxiaoLiu/LRV-Instruction)
- [LAION-GPT4V](https://huggingface.co/datasets/laion/gpt4v-dataset)
- [ShareGPT4V](https://huggingface.co/datasets/Lin-Chen/ShareGPT4V)
- [DocVQA](https://huggingface.co/datasets/pixparse/docvqa-single-page-questions)
- [SynthDoG English](https://huggingface.co/datasets/naver-clova-ix/synthdog-en)
- [ChartQA](https://huggingface.co/datasets/HuggingFaceM4/ChartQA)
- [DVQA](https://huggingface.co/datasets/vikhyatk/dvqa)
- [AI2D subset and image assets](https://huggingface.co/datasets/Lin-Chen/Open-LLaVA-NeXT-mix1M)