README.md

---
configs:
- config_name: default
  data_files:
  - split: train
    path: train-*
  - split: validation
    path: validation-*
dataset_info:
  features:
  - name: texts
    sequence: string
  - name: images
    sequence: image
  - name: metadata
    list:
    - name: document_url
      dtype: string
    - name: unformatted_src
      dtype: string
    - name: src
      dtype: string
    - name: formatted_filename
      dtype: string
    - name: alt_text
      dtype: string
    - name: rendered_width
      dtype: int64
    - name: rendered_height
      dtype: int64
    - name: original_width
      dtype: int64
    - name: original_height
      dtype: int64
    - name: format
      dtype: string
  - name: general_metadata
    struct:
    - name: url
      dtype: string
    - name: warc_filename
      dtype: string
    - name: warc_record_offset
      dtype: int64
    - name: warc_record_length
      dtype: int64
task_categories:
- image-text-to-text
language:
- en
size_categories:
- 1K<n<10K
tags:
- multimodal
- interleaved
- pretraining
---

# Tiny-OBELICS (Debug Dataset)

**Tiny-OBELICS** is a curated, 100% offline-compatible multimodal debug dataset of interleaved
document-image examples drawn from the first shard of `HuggingFaceM4/OBELICS`.

**All image bytes are stored directly inside the Parquet shards** using Hugging Face's native
`Image` feature type. No CDN dependencies, no dead links, no download bubbles at training time.

## Dataset Structure

Parallel-lists layout matching the original OBELICS schema.

| Column | Type | Description |
| :--- | :--- | :--- |
| `texts` | `list[str\|null]` | Text segments interleaved with images |
| `images` | `list[image\|null]` | PIL images (or `None` for text-only slots) |
| `metadata` | `list[struct\|null]` | Per-image metadata (dimensions, src, alt-text, …) |
| `general_metadata` | `struct` | Document-level crawl info (url, warc offset/length) |

## Shard Policy

- Shards are bounded by **raw image byte size** (~25MB per shard), not by row count.
- Any document where one or more images failed to download is discarded entirely.
- Images are converted to RGB and have metadata stripped to ensure clean serialization.
- Original image width and height are always calculated directly from the verified PIL image object.

## Usage

```python
from datasets import load_dataset

# Load train split
ds_train = load_dataset("akzsh/tiny-obelics", split="train")
print(ds_train)

# Load validation split
ds_val = load_dataset("akzsh/tiny-obelics", split="validation")
print(ds_val)

row = ds_train[0]
for text, img in zip(row["texts"], row["images"]):
    if text: print("[Text]", text[:80])
    elif img: print("[Image]", img.size)
```

## Source

Processed subset of [HuggingFaceM4/OBELICS](https://huggingface.co/datasets/HuggingFaceM4/OBELICS).