augustinian-babylm/augustinian_babylm

Multimodal BabyLM — grounded (region, text) table

Unified pool of region/whole-image → text pairs for visual initialisation of text token embeddings. Private because it bundles images from Flickr30k, COCO/RefCOCO, and THINGS, whose licences do not all permit public redistribution.

Counts

• Rows (deduped): 563,118
• Unique images: 88,776
• Per source: {'flickr30k': 309948, 'refcocoplus': 132274, 'refcocog': 94790, 'things': 26106}

Schema (annotations.parquet)

field	meaning
uid	stable row id
source	flickr30k / refcocog / refcocoplus / things
image_uid	content hash of decoded pixels; join key into the image shards
bbox	[x, y, w, h] pixels, or null = whole image
text / text_norm	raw phrase / lowercased-normalised
category	flickr 9-cat / COCO category / THINGS concept-level (nullable)
n_annotations	how many independent annotations back this grounding
sources / provenance	which datasets + original ids contributed

Images live in images/images-NNNNN.parquet (HF Image feature, one row per image_uid); images/manifest.parquet maps each image_uid to its shard.

Load

from huggingface_hub import snapshot_download
from datasets import load_dataset
import pandas as pd, os

d = snapshot_download("bylin/augustinian_babylm", repo_type="dataset")
ann = pd.read_parquet(os.path.join(d, "annotations.parquet"))

# images: a Dataset of (image_uid, image=PIL). Build uid -> row index once.
imgs = load_dataset("parquet",
                    data_files=os.path.join(d, "images", "images-*.parquet"),
                    split="train")
idx = {u: i for i, u in enumerate(imgs["image_uid"])}

# for a row: pil = imgs[idx[row.image_uid]]["image"]; crop with row.bbox