noun-attributes

Images of nouns with visual attributes, with VLM-generated captions and GPT-rated attribute applicability.

Dataset summary

• Total samples: 25,599
• Columns: noun, image, attributes, attribute_rank, vlm_caption

Nouns

• Unique: 245
• Samples per noun — min: 5, max: 255, avg: 104.1

Attributes

• Unique: 56
• Samples per attribute — min: 13, max: 7459, avg: 888.1
• Attributes per sample — min: 1, max: 12, avg: 1.9

Attribute rank 5 (prototypical applicability)

• Samples with ≥1 rank-5 attribute: 10,484
• (noun, attr) pairs with rank 5: 13,354
• Unique nouns with ≥1 rank-5 attribute: 230

VLM captions

Captions are generated by Qwen/Qwen2.5-VL-3B-Instruct. We use the same instruction prompt that was given to MS-COCO annotators:

You are shown an image. Write a single sentence caption. Follow these instructions: Describe all the important parts of the scene. Do not start the sentence with "There is". Do not describe unimportant details. Do not describe things that might have happened in the future or past. Do not describe what a person might say. Do not give people proper names. The sentence should contain at least 8 words.

• Unique captions: 24462
• Words per caption — min: 1, max: 24, avg: 15.1

Steering selection (Pa@10 eval)

When selecting (noun, attribute) pairs for steering, use only pairs that satisfy:

• ≥10 images per (noun, attr) — required for Pa@10 evaluation

• Rank ≥ 3 (GPT applicability rating)

• Steering-eligible pairs: 1676

• Unique nouns (eligible): 239

• Unique attributes (eligible): 40

Code to compute steering-eligible pairs:

from collections import defaultdict
from datasets import load_dataset
import json

ds = load_dataset("nirmalendu01/noun-attributes", split="train")
MIN_IMAGES, MIN_RANK = 10, 3
nouns = [str(r.get("noun") or "").strip() for r in ds]
attrs_per_row = [[str(a).strip() for a in (r.get("attributes") or []) if a] for r in ds]
pair_counts = defaultdict(int)
for i, (noun, attrs) in enumerate(zip(nouns, attrs_per_row)):
    if noun:
        for attr in attrs:
            if attr:
                pair_counts[(noun, attr)] += 1
pair_ratings = {}
for i, r in enumerate(ds):
    ar = r.get("attribute_rank")
    if not ar:
        continue
    try:
        d = json.loads(ar) if isinstance(ar, str) else ar
    except Exception:
        continue
    noun = nouns[i] if i < len(nouns) else ""
    for attr, info in (d.items() if isinstance(d, dict) else []):
        if isinstance(info, dict) and "rating" in info:
            pair_ratings[(noun, attr)] = info["rating"]
eligible = []
for (noun, attr), count in pair_counts.items():
    rating = pair_ratings.get((noun, attr))
    if rating is not None and count >= MIN_IMAGES and rating >= MIN_RANK:
        eligible.append({"noun": noun, "attribute": attr, "count": count, "rank": rating})