---
pretty_name: LLaVA-CoT 48k Reannotated
language:
- en
size_categories:
- 10K<n<100K
tags:
- image
- text
- multimodal
- vision-language
- parquet
- datasets
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: image_path
    dtype: string
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      list:
      - name: type
        dtype: string
      - name: image
        dtype: string
      - name: text
        dtype: string
  - name: relabel_meta
    struct:
    - name: source_sample_id
      dtype: string
    - name: source_line_index
      dtype: int64
    - name: relabel_model
      dtype: string
    - name: relabel_max_tokens
      dtype: int64
    - name: generated_at_utc
      dtype: string
    - name: assistant_turns_rewritten
      dtype: int64
---

# llava-cot-48k-reannotated

This is a Parquet-native repack of the previous JSONL + image-file layout.
Images are embedded directly in the `image` column as bytes, and the original
relative path is preserved in `image_path`.

- rows: 47905
- unique image files in source folder: 38930
- parquet shards: 20
- source: 47,905 filtered rows from LLaVA-CoT-100k_relabel.vl.min10.jsonl
- layout: `data/train-*.parquet`

Each row keeps the original conversation-style supervision and relabel metadata.

```python
from datasets import load_dataset

ds = load_dataset("hao05/llava-cot-48k-reannotated", split="train")
sample = ds[0]
```