mlfoundations/dcvlm_pool_small_annotations

DCVLM-Pool (small) — per-sample annotations

Every filtering annotation we computed for the small data pool of our DataComp-VLM paper: image quality, image–text alignment, language ID, text-quality classifiers, multimodal perplexity, decontamination scores and more — up to 167 fields per sample (180 distinct fields overall), for all 120,940,134 samples across 166 source datasets.

These are the raw annotations, not a filtered dataset. They are the inputs our curation pipeline consumes to build a training mixture: you pick fields, pick thresholds, and get a subset. Releasing them means you can reproduce our filters, or design your own, without recomputing anything on 828 GB of raw data.

The pool these annotate is mlfoundations/dcvlm_pool_small — 12,206 WebDataset shards, 10.3 TB.

Layout

One JSON file per pool shard, under the source dataset's directory — exactly mirroring the pool's own <source>/<shard>.tar layout, so ai2d/000000.json annotates ai2d/000000.tar.

datacomp_1b/000000.json … 004309.json    # 4,310 shards (largest source)
relaion-2b-en-vol-1-unpacked-TEST/…      # 3,015 shards
flanv2/000000.json … 001240.json         # 1,241 shards
ai2d/000000.json                         # 1 shard
…                                        # 166 sources, 12,206 files, 828 GB

Each file is a JSON object mapping the sample's WebDataset __key__ → its annotations. For example:

{
  "481c61f555bfbdecf6d13a73052ef621-0": {
    "length_tokens": 3405, "length_tokens_text_only": 78, "length_images": 1,
    "clip_score_siglip2_b16_224_standard_mean": 0.859,
    "iqa_v2_koniq10k_mean": 0.584,
    "im_height": {"jpg": 365}, "im_width": {"jpg": 512}, "im_entropy_mean": 7.67,
    "language_detection_annotation_lingua_0_lan": "en",
    "text_quality_weborganizer_topic_label": "Travel",
    "text_quality_dolma3_fasttext_score": 0.539,
    "unifilter_score_qwen3_0_6b": 3.125,
    "imdecont_max": 0.226,
    "within_sample_repetition_detection": 10.0
  }
}

The annotations

Family	Fields	What it is
Image–text alignment	48	CLIP-style scores from 4 encoders (clip_vitl14_224, dfn_public, siglip2_b16_224, longclip_vitl14_224) × 4 turn scopes (standard, prev_turn_only, next_turn_only, prev_and_next_turn) × min/mean/max over the sample's images
Image statistics	36	Per-image height, width, aspect ratio, KB, bits-per-pixel, clarity (sharpness), luma, saturation, entropy — raw per-member dict plus min/mean/max
Image quality (IQA)	24	No-reference quality predictors trained on 8 IQA datasets (live, csiq, tid2013, kadid10k, flive, spaq, clive, koniq10k), min/mean/max. A legacy iqa_* (v1) set is also present on some sources; prefer iqa_v2_*
Language ID	14	Top-3 languages + scores from two detectors, lingua and nllb (flattened _0/_1/_2 fields plus the raw _top3 list)
Text quality	10	weborganizer topic + format labels, NVIDIA domain label, two NVIDIA edu-value scores (nemotron4, mixtral), DCLM-baseline and Dolma-3 fastText scores
Decontamination	6	imdecont_* = similarity of each image to our eval-benchmark image index, plus the matched index id (imdecont_image_match); textdecontamination = binary text-side overlap check (1 = clean)
Length / tokens	5	length_chars, length_tokens (full multimodal, incl. image tiles), length_tokens_text_only, length_tokens_trainable_only, length_images
Custom fastText	4	Quality classifiers targeting our val and core evaluation suites
Multimodal perplexity	3	Qwen2.5-VL-3B-Instruct cross-entropy on text-only and image+text, plus their difference as conditional mutual information — how much the image actually helps predict the text
UniFilter	2	Learned data-quality scores from qwen2_5_1_5b and qwen3_0_6b UniFilter models
Heuristic flags	4	detoxify (binary, 1 = passes toxicity check), zero_sequence_filter / digit_sequence_filter (integer counts of degenerate token runs, typically 0–5), within_sample_repetition_detection (integer 0–10 repetition rating within a sample)

Coverage caveats

• The field set varies by source (106–167 fields), because some annotators only apply to some data. It is uniform within a file, so reading one record tells you that source's schema.
• Missing ≠ failed. textdecontamination is only present on text-bearing sources, detoxify is absent on some web-scale captioning sources. Treat an absent field as "not applicable".
• Image fields are null for text-only samples. imdecont_* instead uses sentinels: -1.0 when the sample has no images at all, and -1.7e38 (float32 min, with imdecont_image_match == -1) when an image returned no index match. Both mean "uncontaminated" and both pass a < threshold test — so compare raw values directly and don't normalize or take absolute values.

Usage

Grab one source (each file is 5–90 MB, so pull only what you need):

from huggingface_hub import hf_hub_download
import json

p = hf_hub_download("mlfoundations/dcvlm_pool_small_annotations",
                    "ai2d/000000.json", repo_type="dataset")
ann = json.load(open(p))          # {sample_key: {field: value}}

Or a whole source directory:

from huggingface_hub import snapshot_download
snapshot_download("mlfoundations/dcvlm_pool_small_annotations", repo_type="dataset",
                  allow_patterns="chartqa/*")

Then select keys and keep only those samples when reading the matching pool shard:

# The decontamination applied to every source in our released mixtures:
keep = {
    k for k, a in ann.items()
    if a["imdecont_max"] < 0.75                       # no near-duplicate of an eval image
    and (a.get("textdecontamination", 1) > 0.0)       # no eval-text overlap (absent => passes)
}

# Layer on whatever else you want, e.g. a quality + alignment + language filter:
keep = {
    k for k in keep
    if (ann[k].get("iqa_v2_koniq10k_mean") or 0) > 0.5
    and (ann[k].get("clip_score_clip_vitl14_224_standard_mean") or 0) > 25
    and ann[k]["language_detection_annotation_lingua_0_lan"] == "en"
}

⚠️ Alignment scores are not on a shared scale — inspect each encoder's own distribution before picking a threshold. On datacomp_1b, clip_vitl14_224 and longclip_vitl14_224 run ~20–35, dfn_public ~0–5, and siglip2_b16_224 is an unbounded logit spanning roughly −17 to 20.

The __key__ values match the pool shard exactly, so filtering is a set-membership test while streaming:

import webdataset as wds
ds = wds.WebDataset("ai2d/000000.tar").select(lambda s: s["__key__"] in keep)

Citation

@article{farina2026datacomp,
  title={DataComp-VLM: Improved Open Datasets for Vision-Language Models},
  author={Farina, Matteo and Udandarao, Vishaal and Nguyen, Thao and Kuzucu, Selim and B{\"o}ther, Maximilian and Hochlehnert, Andreas and Ghosh, Adhiraj and Nezhurina, Marianna and Roth, Karsten and Struber, Joschka and others},
  journal={arXiv preprint arXiv:2606.28551},
  year={2026}
}