Relationship to pages:
comix-v0_1-books extends comix-v0_1-pages.
Every book record is built from the pages dataset and provides:

book-level metadata (placeholder in v0.1)

ordered list of pages and shard locations

hooks for future segments/characters annotations.

What's Included

Each sample corresponds to one book and contains a single JSON file:

{book_id}.json - book metadata and page references

Book JSON Schema (v0.1)

{
  "book_id": "c00004",
  "book_metadata": {
    "series_title": null,
    "issue_number": null,
    "publication_date": null,
    "publisher": null,
    "total_pages": 68,
    "license_status": "Public Domain",
    "digital_source": "Digital Comic Museum"
  },
  "pages": [
    {
      "page_number": 0,
      "page_id": "c00004_p000",
      "tar_file": "pages-train-00000.tar",
      "has_segmentation": true
    }
    // ...
  ],
  "segments": [],     // v2+: story segments (to be added)
  "characters": []    // v2+: character bank (to be added)
}

Primordial status (v0.1):

book_metadata fields are mostly empty/null for now.

segments and characters are placeholders for future versions.

Future releases (v2+) will fill bibliographic information, story segments, summaries and richer character annotations.

Data Splits

Split	Books
Train	19158
Validation	2
Test	5
Total	19165

Split Strategy

Splits are exactly aligned with the pages dataset:

Splits are defined per book, not per page.
Each book_id is assigned to train / validation / test using an MD5 hash-based mapping consistent with the CoMix benchmark (C100 + DCM) as defined in: https://github.com/emanuelevivoli/CoMix
All pages of a given book are in the same split in both datasets.
Books not present in the benchmark split lists fall back to train.

Sharding & Relationship with `pages`

Books are created after the pages dataset:

Build comix-v0_1-pages from groups group_00-group_14.
For each group, read the page JSONs and shard metadata.
Group pages by book_id and split.
For each book, create a JSON file referencing:
- ordered list of page_ids
- the tar_file of the corresponding page in the pages dataset
- whether segmentation is available (has_segmentation)

Like the pages dataset, books are sharded as:

books-train-XXXXX.tar
books-validation-XXXXX.tar
books-test-XXXXX.tar

with the same 5-digit index XXXXX corresponding to the original source tar.

Tiny versions (*-books-tiny) are built from a reduced number of tars per group and aggregated in the same way.

Quick Start (Hugging Face `datasets`)

from datasets import load_dataset
import json

# Load books dataset (streaming recommended for large-scale use)
books = load_dataset(
    "emanuelevivoli/comix-v0_1-books",
    split="train",
    streaming=True,
)

for book in books:
    book_data = json.loads(book["json"])

    book_id = book_data["book_id"]
    total_pages = book_data["book_metadata"]["total_pages"]

    # Iterate through page references
    for page_ref in book_data["pages"]:
        page_id = page_ref["page_id"]
        page_number = page_ref["page_number"]
        tar_file = page_ref["tar_file"]
        has_seg = page_ref["has_segmentation"]

    print(f"Book {book_id}: {total_pages} pages (train split)")

Example: Join Books with Pages

from datasets import load_dataset
import json

pages = load_dataset("emanuelevivoli/comix-v0_1-pages", split="train")
books = load_dataset("emanuelevivoli/comix-v0_1-books", split="train")

# Build a simple index from page_id to page sample
page_index = {p["json"]["page_id"]: p for p in pages}

book = books[0]
book_data = json.loads(book["json"])

book_pages = []
for page_ref in book_data["pages"]:
    pid = page_ref["page_id"]
    if pid in page_index:
        book_pages.append(page_index[pid])

print(f"Book {book_data['book_id']}: {len(book_pages)} pages loaded")

Use Cases

Book-level analysis - model complete story arcs/book structure
Page grouping - group pages into sequences for autoregressive tasks
Metadata enrichment - attach external metadata at book level
Benchmark alignment - leverage CoMix splits in downstream tasks
Multimodal pipelines - use book metadata + page images for VQA, story generation, summarisation, etc.

Because books extends pages, this is the recommended way to work at “comic book” granularity while keeping the heavy pixel data in a shared page-level dataset.

Known Limitations (v0.1)

Primordial metadata:
- Most bibliographic fields (series_title, issue_number, publication_date, publisher, etc.) are empty/null.
- segments and characters are empty lists for now.
Dependency on pages:
- This dataset does not contain images. All visual information must be loaded from emanuelevivoli/comix-v0_1-pages.
- Page references (tar_file, page_id) assume the v0.1 layout of the pages dataset and may change in future major versions.
Noisy / missing lower-level annotations:
- Any noisiness or missing data in the pages dataset (detections, captions, segmentations, empty seg.npz files) naturally propagates here.

If you run into issues, please get in touch: emanuele [dot] vivoli [at] unifi [dot] it.

Processing Pipeline (High-Level)

Group pages by book using book_id from the pages JSON.
Compute split for each book using MD5 hash and CoMix split lists.
Create book JSON with:
- placeholder book_metadata
- ordered pages references (with tar_file and has_segmentation)
- empty segments and characters for future v2+ annotations.
Export WebDataset shards mirroring the pages shard indices.

Tiny Subset

For experimenting with book-level code:

Books tiny: emanuelevivoli/comix-v0_1-books-tiny
- Very small number of shards, aligned with the tiny pages subset.

Citation

@dataset{comix_v0_1_books_2025,
  title   = {Comic Books Dataset v0.1 - Books},
  author  = {Emanuele Vivoli},
  year    = {2025},
  note    = {Book-level metadata built on top of the comix-v0.1 pages dataset},
  url     = {https://huggingface.co/datasets/emanuelevivoli/comix-v0_1-books}
}

License

Dataset: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).
Underlying comic scans are from public-domain sources.
If you extend these metadata (e.g. enrich bibliographic fields, add segments/characters annotations), please share your work with the community under a compatible share-alike license.