tsmithv11's picture
Credit the original Stack Overflow screenshot excerpt (#2)
bad61ee
|
Raw History Blame Contribute Delete
6 kB
metadata
pretty_name: Hugging Face Course Documentation Images (Lance)
license: apache-2.0
language:
  - en
tags:
  - lance
  - course
  - documentation
  - image-retrieval
task_categories:
  - image-to-text
configs:
  - config_name: default
    data_files:
      - split: test
        path: viewer/test.parquet
  - config_name: catalog
    data_files:
      - split: assets
        path: viewer/assets.parquet

Hugging Face Course Documentation Images in Lance

The Hugging Face LLM Course uses these diagrams, screenshots, illustrations, and animations in its lessons. This conversion preserves all 135 files and their original paths, including 132 images. The upstream Hub viewer has one two-image test split, which is preserved in Lance and as the exact original Parquet preview. The added image search table covers the complete asset collection. Original dataset · LLM Course

Path Rows Contents
lance/original-files.lance 135 Every original file with bytes, relative path, size, MIME type, and hashes
lance/images.lance 132 Image paths, metadata, and search embeddings
lance/test.lance 2 Original Hub viewer test image rows
viewer/test.parquet 2 Original Hub viewer Parquet
viewer/assets.parquet 132 Thumbnails for browsing the full image collection

The "catalog" preview uses SVG renders and the first frame of each GIF. The files table retains the original SVGs, complete animations, and exact source bytes. Source tree inventory · Release metadata

Search images with text

The release generates 1,152-dimensional float32 image embeddings with the visual encoder of Google SigLIP 2 google/siglip2-so400m-patch16-384 at revision dd658faac399427308559e2c3ac1e99cbe43845d. Images were composited onto white where transparent. SVGs use rasterized previews and GIFs use their first frame. The example uses the model's paired text encoder, normalizes both modalities to unit length, and prints five source paths with cosine similarity, calculated as one minus Lance's cosine distance. Pinned SigLIP 2 checkpoint

Install the dependencies with python -m pip install "pylance==14.0.3" "transformers==5.23.0" torch pillow huggingface_hub.

import lance
import torch
from huggingface_hub import HfApi
from transformers import AutoModel, AutoProcessor

repo = "lance-format/documentation-images-lance"
revision = HfApi(token=False).dataset_info(repo, token=False).sha
uri = f"hf://datasets/{repo}/lance"
options = {"hf_revision": revision}
images = lance.dataset(f"{uri}/images.lance", storage_options=options)

checkpoint = "google/siglip2-so400m-patch16-384"
model_revision = "dd658faac399427308559e2c3ac1e99cbe43845d"
encoder = AutoModel.from_pretrained(checkpoint, revision=model_revision).eval()
processor = AutoProcessor.from_pretrained(checkpoint, revision=model_revision)
query = "a diagram of a transformer model with an encoder and a decoder"
inputs = processor(text=[query], padding="max_length", truncation=True, return_tensors="pt")
with torch.inference_mode():
    pooled = encoder.get_text_features(**inputs).pooler_output
    vector = torch.nn.functional.normalize(pooled.float(), dim=-1)[0].numpy()

hits = images.to_table(
    columns=["source_path", "content_type", "_distance"],
    nearest={"column": "embedding", "q": vector, "k": 5, "metric": "cosine"},
).to_pylist()
for hit in hits:
    print(f"{1 - hit['_distance']:.4f}  {hit['source_path']}  {hit['content_type']}")

Retrieve an original asset

from pathlib import Path

import lance
from huggingface_hub import HfApi

repo = "lance-format/documentation-images-lance"
revision = HfApi(token=False).dataset_info(repo, token=False).sha
files = lance.dataset(
    f"hf://datasets/{repo}/lance/original-files.lance",
    storage_options={"hf_revision": revision},
)
asset, = files.to_table(
    columns=["original_bytes"],
    filter="source_path = 'en/chapter1/transformers_architecture.png'",
    limit=1,
).to_pylist()
Path("transformers_architecture.png").write_bytes(asset["original_bytes"])

Source and license

Original assets are credited to Hugging Face and the LLM Course contributors. The dataset card declares Apache 2.0, and the course project includes the Apache 2.0 license. Changes in this release consist of added Lance tables, image embeddings, thumbnails, and file metadata. Image trademarks and other embedded notices remain visible in the originals. Original card · Apache 2.0 license · Course source

Credits for screenshots of published posts

  • en/chapter8/stack-overflow.png reproduces an answer by cronoik, posted on March 14, 2021. The embedded answer remains under CC BY-SA 4.0. The Lance files table preserves the exact course screenshot, and the catalog preview displays a resized copy. Original answer · CC BY-SA 4.0 · Stack Overflow licensing
  • en/chapter5/datasets-issues-comment.png shows yjernite and patrickvonplaten discussing a Common Voice update. en/chapter5/datasets-issues-single.png shows lewtun's Wikipedia streaming issue. en/chapter5/datasets-issues.png shows the issue list. Common Voice issue · Wikipedia streaming issue · Hugging Face Datasets issue list