mrdbourke/food-or-drink-10m

Food or Drink 10M

A balanced binary classification dataset for detecting food or drink content in image captions.

Overview

	Count
Total rows	1,572,911
Train split	1,415,619 (90%)
Test split	157,292 (10%)
Food/drink rows	785,980 (50.0%)
Not food/drink rows	786,931 (50.0%)

How it was made

1. Source: UCSC-VLAA/Recap-DataComp-1B (condition_diverse_topk subset)
2. Teacher model: MoritzLaurer/ModernBERT-large-zeroshot-v2.0 — a 400M parameter zero-shot NLI classifier
3. Classification: 10M captions streamed and classified as food_or_drink / not_food_or_drink using zero-shot NLI with candidate labels ["food or drink", "not food or drink"]
4. Balancing: All food/drink rows saved (100%), not-food/drink rows dynamically sampled to match — resulting in a ~50/50 balanced dataset

Labels

• food_or_drink — caption describes food, beverages, meals, ingredients, drinks, or food/drink items
• not_food_or_drink — caption describes anything else (objects, scenes, people, animals, etc.)

Fields

Field	Description
url	Original image URL from DataComp-1B
re_caption	AI-generated re-caption (detailed, descriptive)
org_caption	Original caption (often noisy alt-text)
re_caption_condition_diverse_topk	Condition-diverse re-caption variant
label	Classification label: food_or_drink or not_food_or_drink
score	Teacher model confidence score (0.5–1.0)
sha256	Image content hash
key	Row key from source dataset
re_clip_score / org_clip_score	CLIP alignment scores
re_length / org_length	Caption token lengths
re_gpt4v_score / org_gpt4v_score	GPT-4V quality scores
re_condition_length	Condition caption token length

Intended use

This dataset is designed for:

• Knowledge distillation: Fine-tuning smaller encoders (e.g., Ettin-encoder-150m) to replicate the teacher model's classification, then running the fine-tuned model on the full 1B-row dataset
• Food/drink content filtering: Filtering large-scale image-text datasets for food and drink related content
• Caption classification research: Studying food/drink detection in image captions

Caption types

The dataset contains multiple caption styles per image, useful for training robust classifiers:

• re_caption: Clean, AI-generated descriptions (e.g., "A glass of red wine next to a cheese board on a wooden table")
• org_caption: Original noisy alt-text (e.g., "IMG_2847 wine night")
• re_caption_condition_diverse_topk: Longer, more detailed AI captions

Example samples

Food or drink

label	score	re_caption	org_caption
food_or_drink	0.9988	A glass of minty chocolate latte with a straw is placed on a yellow surface next to a bag of coffee beans.	Mint Chocolate Coffee Frappe Recipe
food_or_drink	0.9978	A plate of food with a mix of lettuce, meat, and sauce. The plate is on a dining table with a glass of orange juice and a bowl of salad.	Summer lunch spread with fresh OJ

Not food or drink

label	score	re_caption	org_caption
not_food_or_drink	0.9998	A row of identical figures in black suits and ties is standing in a line against a white background.	O 6-Car Flat End Offset Hopper Car Set, B&O 727022
not_food_or_drink	0.9997	A white wainscoting panel with a decorative molding at the top and a plain lower section is attached to a wall.	How to Install Board and Batten Wainscoting (White Painted Square over Rectangle Pattern)

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("mrdbourke/food-or-drink-10m")

# Access splits
train = ds["train"]
test = ds["test"]

print(f"Train: {len(train):,} rows")
print(f"Test:  {len(test):,} rows")

# Look at a sample
print(train[0])

# Filter to food/drink only
food_only = train.filter(lambda x: x["label"] == 0)  # 0 = food_or_drink
print(f"Food/drink rows: {len(food_only):,}")

# Stream instead of downloading
ds_stream = load_dataset("mrdbourke/food-or-drink-10m", split="train", streaming=True)
for row in ds_stream:
    print(row["re_caption"], row["label"])
    break

Confidence scores

The score field contains the teacher model's confidence. Higher scores indicate more certain classifications:

Score range	Meaning
0.95–1.0	Very confident
0.80–0.95	Confident
0.60–0.80	Moderate confidence
0.50–0.60	Low confidence (borderline)

License

Apache 2.0 — same as the source dataset and teacher model.