LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models

“LAION-5B: An open large-scale dataset for training next generation image-text models” is a paper by Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk and Jenia Jitsev, published in the Datasets and Benchmarks Track of NeurIPS 2022. The authors came from LAION, which the datasheet describes as a non-profit research organization (Q2), together with UC Berkeley, the University of Washington, TU Darmstadt, TU Munich, the Jülich Supercomputing Centre, Stability AI, EleutherAI and others. The paper presented LAION-5B, 5.85 billion image URLs paired with their web alt-text and kept only when a CLIP model judged image and text similar, and argued that an open dataset of this size would let researchers outside a few industrial labs train and audit image–text models. To show that the data were usable, the authors trained CLIP models on its subsets that came close to OpenAI’s and fine-tuned a text-to-image model on it. The paper received one of two Outstanding Datasets and Benchmarks Paper awards at NeurIPS 2022. This article covers the publication; the LAION article covers the datasets as they are today, their licence, the audits of their content, the December 2023 withdrawal and the 2024 re-release.

This article is based on the paper and the supplemental material on the NeurIPS proceedings site. The supplement holds Appendices A–G (a datasheet, setup and curation details, statistics, CLIP and generative-model experiments, and a safety and ethics discussion), the author contributions and the NeurIPS checklist. The arXiv version (2210.08402v1, 16 October 2022, the only arXiv version as of October 2026) has the same appendices and author contributions, word for word apart from layout, but not the checklist, and the same main text except for one sentence in §5.2.1 that appears only in the proceedings version (see Results). Section, table and figure numbers below are the same in both.

Problem

The paper opens with CLIP, whose contrastive pretraining on 400 million image–text pairs had raised zero-shot ImageNet top-1 accuracy from the previous 11.5% to 76.2%, and with DALL·E and its successors in text-to-image generation (§1). All of these depended on datasets of hundreds of millions to billions of pairs, such as BASIC’s 6.6 billion, and none of those datasets was public; OpenAI released the CLIP weights, but later groups released neither data nor models. The authors argue that research in this area had therefore become concentrated in a few industrial labs, which limited transparency and slowed progress (§1).

Table 1 sets out the gap in English pairs. The public datasets were MS-COCO (330K), CC3M (3M), Visual Genome (5.4M), WIT (5.5M), CC12M and RedCaps (12M each) and YFCC100M (100M, though footnote 2 notes that it is unclear how well its text matches its images and that CLIP’s curation kept 15M of them). The private ones were CLIP’s WIT (400M), ALIGN (1.8B) and BASIC (6.6B). §2 reviews how the public datasets were built: COCO and Visual Genome by human annotation, which kept them small; YFCC100M from Flickr user metadata; Conceptual Captions from web alt-text with additional cleaning, which CC12M relaxed to grow; ALT200M and ALIGN from alt-text at larger scale; and RedCaps from Reddit captions. For other languages the largest dataset the authors knew of had about 36 million samples, from Wikipedia-based Image Text.

The authors’ own earlier release, LAION-400M (Schuhmann et al., 2021), had already offered 400 million English pairs filtered with CLIP at a cosine similarity of 0.3. LAION-5B was meant to reach the scale of the largest private datasets.

Contribution

• A dataset of 5.85 billion pairs in three subsets: 2.32 billion English pairs (LAION-2B-en), 2.26 billion in more than 100 other languages, and 1.27 billion with no language detected with confidence (§1, §4). The release consists of metadata: URLs, alt-text and scores, not images.
• Tools and derived data: the crawling code, the img2dataset downloader, CLIP ViT-L/14 image embeddings, k-nearest-neighbour indices and a public search interface (§3.1, §7, Appendices B and C).
• Safety and quality tags: scores from an NSFW classifier and a watermark detector for every image, with offensive content tagged rather than removed (§3.2, §4).
• Validation by training: OpenCLIP reproductions of CLIP ViT-B/32, B/16 and L/14 on LAION-400M and LAION-2B-en, whose ViT-L/14 models the authors call the first open-source reproductions of OpenAI’s largest CLIP models (§1, §5.2), and a GLIDE model fine-tuned on LAION-5B (§5.3).
• Documentation: a datasheet answering the standard questions on motivation, composition, collection, uses, distribution and maintenance (Appendix A), and discussions of technical limitations (§6) and safety and ethics (§7, Appendix G).

The authors present the release as a first step and not, in their words, “a finished data product” (§1). They strongly recommend that LAION-5B in its current form be used only for academic research, and advise against deployed applications without careful study of the behaviour and biases of models trained on it (§1, in bold in the paper; §7).

Method

Collection

The LAION article describes the pipeline (Figure 2 of the paper) step by step; in outline, workers parsed the IMG tags with alt-text from Common Crawl’s WAT metadata files, ran the CLD3 language detector on the alt-text, downloaded the images, dropped pairs with less than 5 characters of text or 5 KB of image data as well as images judged potentially malicious, large or redundant, and kept a pair only if its CLIP image and text embeddings had a cosine similarity of at least 0.28 for English (OpenAI’s ViT-B/32) or 0.26 for everything else (a multilingual ViT-B/32 by Carlsson et al.). This removed about 90% of some 50 billion candidates (§3.1). The paper adds engineering details that the dataset article omits:

• A PostgreSQL server held about 500 million image URLs at any time and handed out work in chunks of 10,000 links. Each of roughly 300 workers was a small cloud node (2 vCPUs, 1 GB of RAM, 10 Mbps), downloading asynchronously with the Python libraries Trio and Asks; a worker took 10–15 minutes per chunk (§3.1).
• Larger CLIP models had not been released when the work began, so ViT-B/32 was used throughout for consistency (§3.1).
• The datasheet says the data were filtered from September 2021 to January 2022 (Q26), that no crowdworkers were used, and that open-source contributors did the work unpaid (Q25). Hugging Face and Stability AI sponsored it (Q3), and compute for the CLIP experiments came from the JUWELS Booster supercomputer at Jülich (Acknowledgments).

The paper does not say how duplicates were found. The LAION-5B announcement says a Bloom filter on the URL removed them (Beaumont, 2022), and the author contributions credit Beaumont with a PySpark deduplication job.

After filtering

Appendix B lists four steps that turn the filtered URL list into the released resource, starting from a fresh download of the images:

1. Downloading with img2dataset, which the authors measured at 100 million images in 20 hours on one machine (1 Gbps, 32 GB of RAM, a 16-core i7), and in a distributed mode at all 5 billion samples in a week on 10 nodes (C.1).
2. Embedding every image with CLIP ViT-L/14, which took a week on 32 NVIDIA A100 GPUs; the larger model processes 312 samples per second per GPU against 1,800 for ViT-B/32 (C.2).
3. Indexing the 9 TB of embeddings with autofaiss into a product-quantized (PQ128) k-nearest-neighbour index (C.3). The 800 GB index and the metadata were memory-mapped from disk to serve the public search interface, which takes text or an image as the query (C.4, §7); the approximate nearest-neighbour search article explains the technique.
4. Tagging each sample with NSFW and watermark scores computed from the CLIP embeddings.

Safety tagging

The paper’s stated policy is to tag rather than remove. Apart from illegal content, which the authors say they filtered using CLIP embeddings without describing how, potentially offensive samples stay in the dataset with tags, so that users can decide per task and researchers can study curation; the authors ask model developers to state in their model cards which subsets and tags they used (§3.2). They warn that the classifiers are imperfect and that removing every tagged sample should not be taken to produce a truly safe subset ready for production (§3.2; datasheet Q16).

Two classifiers are named, both described as working on CLIP embeddings (§3.2): Q16, a published method for documenting inappropriate content, whose most frequent concepts in LAION-2B-en are shown as a word cloud (Appendix C.6, Figure 5), and LAION’s own NSFW classifier. Appendix C.5 describes the NSFW classifier’s training data: more than 682,000 images retrieved from LAION-400M near keywords for five classes, with Wikiart and Danbooru images added, giving 369,507 “neutral”, 207,969 “porn”, 39,026 “drawing”, 37,914 “sexy” and 28,134 “hentai” images. On a manually checked test set of 1,000 images per class, treating drawing and neutral as safe, the classifier caught 96.45% of unsafe images and wrongly flagged 7.96% of safe ones. These numbers are for a fine-tuned EfficientNetV2-B02 image classifier, which C.5 says the authors trained first before moving to a small MLP on CLIP ViT-L/14 embeddings; the appendix gives no accuracy for the MLP. Only the LAION-5B announcement does (0.960 and 0.961 on a balanced set of 3,000 samples, with ViT-B/32 and ViT-L/14 embeddings; Beaumont, 2022). For the watermark and safety tags computed on CLIP embeddings, the paper says only that the training sets were enlarged with images retrieved from the index and that the test sets were manually annotated, with the watermark criteria illustrated in Figure 6 (C.7).

The paper credits this tagging to criticism of the earlier release. After LAION-400M, other groups had used the search tools to examine it, citing Birhane, Prabhu and Kahembwe (2021), an arXiv preprint; the improved inappropriate-content tagging and the watermark filter of LAION-5B were the response (§7).

Training CLIP on the data

The authors adapted OpenCLIP for distributed training on JUWELS Booster and trained ViT-B/32, ViT-B/16 and ViT-L/14 with the hyperparameters of the original CLIP models, plus a “ViT-B/16+” of their own: 240-pixel inputs, wider image and text towers and 208 million parameters, against 150 million for ViT-B/16 (§5.2, Table 3). Training used the InfoNCE loss with AdamW ($\beta_1 = 0.9$, $\beta_2 = 0.98$, weight decay 0.2), linear warmup and cosine decay, on 112 to 400 A100 GPUs with global batches of 32,768 to 46,592 (Appendix E.1, Table 4). Table 4 lists five runs: ViT-B/32 on LAION-400M (32 epochs, 36 hours on 128 GPUs), ViT-B/32 on LAION-2B (16 epochs, 210 hours on 112 GPUs), ViT-B/16 and B/16+ on LAION-400M (61 hours each), and ViT-L/14 on LAION-400M (88 hours on 400 GPUs).

Appendix E.2 explains how the contrastive loss was computed across GPUs. Each worker gathers the embeddings from all others and uses them as negatives, but instead of materializing the full $N \times N$ similarity matrix for a global batch of $N$ pairs, each worker builds only its $n \times N$ slices, where $n$ is its local batch, cutting memory from $O(N^2)$ to $O(nN)$, as in OpenAI’s implementation. The paper puts the naive matrix for $N = 32{,}768$ at 8 GB in float32; one such matrix holds $32{,}768^2 \times 4$ bytes, exactly 4 GiB (about 4.3 GB), so the 8 GB figure matches two of them, one per direction of the loss, as in the two $n \times N$ matrices of the sharded version (our arithmetic; the paper does not show its own). The contrastive learning and CLIP articles explain the loss itself.

Evaluation followed CLIP: each class embedding is the mean of the text embeddings of several prompts collected from earlier work, and an image is assigned to the class with the highest cosine similarity, scored by top-1 accuracy (§5.2.1). Besides ImageNet and four distribution-shift test sets, the authors assembled VTAB+, a superset of VTAB with 35 tasks ranging from CIFAR, Stanford Cars and SUN397 to diabetic retinopathy, satellite images, CLEVR counting and KITTI distances (Appendix E.3, Table 5), and averaged top-1 accuracy over the tasks. A footnote justifies the plain average: VTAB’s authors had found that other aggregation methods correlate highly in rank with it, and averaging per-task ranks gave a similar ordering here (§5.2.1, footnote 10). The prompts and class names were released in the CLIP_benchmark repository (Appendix E.3).

Results

CLIP reproduction

Table 2 compares the LAION-trained models with OpenAI’s models trained on WIT (zero-shot top-1, %; selected columns, with the paper’s full set being ImageNet, ImageNet-v2, ImageNet-R, ImageNet-Sketch, ObjectNet and VTAB+):

Model	Training data	ImageNet	ImageNet-Sketch	ObjectNet	VTAB+ (35 tasks)
ViT-B/32	CLIP WIT	63.3	42.3	44.2	45.4
ViT-B/32	LAION-400M	62.9	49.4	43.9	45.6
ViT-B/32	LAION-2B-en	65.7	52.9	48.7	47.9
ViT-B/16	CLIP WIT	68.3	48.2	55.3	47.5
ViT-B/16	LAION-400M	67.0	52.4	51.5	48.3
ViT-B/16+	LAION-400M	69.2	54.4	53.9	49.2
ViT-L/14	CLIP WIT	75.6	59.6	69.0	55.7
ViT-L/14	LAION-400M	72.8	59.6	59.9	51.8
ViT-L/14	LAION-2B-en	75.2	63.3	65.5	54.6

The authors read this as LAION-400M matching CLIP on ImageNet and in robustness overall, with the largest differences on ObjectNet (WIT better) and ImageNet-Sketch (LAION better), and LAION-2B-en improving on LAION-400M in every column for both model sizes it was used with (§5.2.1). At ViT-L/14 the gap on LAION-400M is 2.8 points on ImageNet and 9.1 on ObjectNet, so in our reading “matched well” holds best for the smaller models. The proceedings version adds a sentence absent from arXiv v1: VTAB+ performance is similar except at ViT-L/14, where WIT has the advantage.

Two analyses address scale. Figure 4 plots ImageNet and VTAB+ accuracy against total training compute (giga multiply–accumulates per sample times samples seen) for the six LAION models, with a straight-line fit on log–log axes; the authors read it as accuracy rising with model, data and compute scale, and leave open whether the trend saturates. Table 6 and Figure 12 hold compute fixed and vary the data: ViT-B/32 trained on LAION-80M, LAION-400M and LAION-2B-en for 3, 13 and 34 billion samples seen, and ViT-L/14 on LAION-400M and LAION-2B-en for 13 and 34 billion. With small budgets the larger dataset does not help (ViT-B/32 at 3 billion samples: 58.73% on LAION-400M, 57.60% on LAION-2B-en), but at 34 billion it does (ViT-B/32: 64.07% against 65.50%; ViT-L/14: 73.90% against 75.40%). The authors conclude that dataset scale matters given sufficient compute (§5.2.1, Appendix E.3).

The appendix adds results beyond Table 2:

• All 35 VTAB+ tasks (Table 7). The appendix notes that LAION-2B-en is not better than LAION-400M on some specialized domains (diabetic retinopathy, PatchCamelyon) and structured tasks. ImageNet-A shows the largest gaps in favour of WIT, 24.3 points at ViT-L/14, which the authors cannot explain; they note that its images were selected to defeat one ResNet-50 and question whether gains on it generalize. On diabetic retinopathy, accuracy swings from 3.3% to 73.3% across WIT models; the class imbalance (74% in the majority class) makes accuracy misleading there, and with mean per-class recall the largest WIT–LAION difference falls to 2.1 points, the best model reaching only 25.4%.
• Few-shot linear probes on ImageNet and six other datasets (Figures 10 and 11), where LAION and WIT models transfer similarly.
• Zero-shot retrieval on the Flickr30k 1K and COCO 5K test sets (Tables 8 and 9), scored by Recall@K. LAION models are better on COCO for every metric and model size. The best COCO results, from ViT-L/14 on LAION-400M, are R@1 of 59.3% for image-to-text and 42.0% for text-to-image retrieval, against 58.4% and 37.8%, which the paper gives as the best results reported by Radford et al. (2021).

Generative models

The generative experiments are qualitative. OpenAI had released GLIDE checkpoints trained on data filtered to remove people and hate symbols, so they could not draw humans; the authors fine-tuned the 64-pixel base model on LAION data to restore that ability, keeping OpenAI’s upsampler (§5.3, Appendix F.1). LAIONIDE-v1 was trained on a single consumer GPU with a batch size of one. LAIONIDE-v2 saw about 25 million LAION-2B pairs after filters on the NSFW tag, language, aspect ratio, a minimum side of 256 pixels and a list of slurs, which the authors call far from comprehensive. LAIONIDE-v3 was fine-tuned on about a million pairs mixed from COCO, Visual Genome, Open Images and LAION-5B to reduce the watermarks that v2 produced. Comparisons with GLIDE are offered as sample images (Figure 13) and an external web demo, with no quantitative metric.

Appendix F.2 lists the training stages of Stable Diffusion on LAION-2B-en, laion-high-resolution and laion-improved-aesthetics; the Stable Diffusion article gives them with the naming differences between official sources.

Uses reported in the paper

§5.1 surveys early uses of LAION data: LAION-5B and LAION-400M as data sources for face representation learning and invasive-species mitigation; LAION’s own LAION-High-Resolution (170 million pairs, for super-resolution) and LAION-Aesthetics (120 million, selected by a linear model on CLIP embeddings); BLIP, MAGMA and an enhanced CLIP trained on LAION subsets; and text-to-image models. Two of these attributions do not match the cited papers:

• §5.1 says the latent diffusion models of Rombach et al. were trained on a subset of LAION-5B, but the datasheet (Q39) says LAION-400M, as does the CVPR paper itself for its text-to-image model (Rombach et al., 2022, §4.3.1; see the latent diffusion paper).
• §5.1 says Imagen was trained on a subset of LAION-400M. The Imagen paper says it combined internal datasets of about 460 million pairs with the publicly available LAION-400M (Saharia et al., 2022, §4.1), so its COCO FID of 7.27, which §5.1 quotes, is not a result for LAION data alone.

Safety and Ethics Statements

§7 makes the case for openness: because most large models and their datasets are closed, outsiders cannot study why models behave as they do, and an open dataset lets the community audit it, measure how uncurated data affect model biases and develop curation methods. The authors call LAION-5B a starting point rather than an endpoint, a research artefact for academic use only, and invite users to report undetected content to a dedicated GitHub repository (§7). A pointer to Appendix G.1 handles privacy.

The datasheet and Appendix G state the authors’ position on people in the data. The LAION article summarizes the answers on consent, notification, identification, removal and ethics review (Q19, Q27, Q30–Q32, Q56); the paper adds:

• Who and what is in it. People appear in the images and text but are not the focus; no subpopulation attributes are provided, though they may sometimes be inferred (Q17–Q18). The data were collected on the assumption that what was public was meant to be shared, yet confidential material may be among it (Q15), and the dataset contains sensitive content (Q20) as well as sexually explicit, racist and abusive content, kept for safety research (Q16).
• Restrictions on use. Given the known biases, no model should be put into production using the dataset as it is, and it should not be used to aid military or surveillance tasks (Q43) or for decisions about people (Q42). Citing Birhane et al., the authors say they authorize the dataset for purely academic work and strongly advise against use in end products (Q33). The LAION article compares these recommendations with the CC BY 4.0 licence of the metadata (Q48–Q49).
• Privacy mitigations. Appendix G.1 argues that the search tool lets the public see, for the first time, what a typical crawled training set contains, and names face obfuscation and face generation as mitigations on which, it says, the LAION community was working. Appendix G.2 notes that CLIP’s own training data are unknown, so the filter’s bias cannot be assessed, and that cosine similarity is a heuristic that Birhane et al. had shown can rank a malignant description above a benign one.
• Confidence in the tags. Appendix C.7 says the tags filter out almost all unsafe results in the search interface, giving confidence that a generative model trained on the data will not produce unexpectedly unsafe images.

Impact

NeurIPS named the paper one of two Outstanding Datasets and Benchmarks Papers of 2022, alongside MineDojo (NeurIPS Blog, 21 November 2022). Much of its influence came through the models trained on its data. The trained CLIP models were released in the OpenCLIP repository (§5.2), and Cherti et al. (2023), several of them authors of this paper, fitted scaling laws for CLIP on LAION data and trained a ViT-H/14 on LAION-2B that reached 78.0% zero-shot ImageNet accuracy (their Table 1, which also lists a ViT-L/14 trained on LAION-2B at 75.2%, the value in this paper’s Table 2). Stable Diffusion was trained on LAION-5B subsets (Appendix F.2). The released embeddings and indices also made LAION a pool from which narrower datasets were selected, as the paper’s own aesthetic and high-resolution subsets show. The vision–language pretraining and text-to-image generation articles survey the models that followed.

In our reading, the paper’s argument that openness enables auditing was borne out, though the audits found more than the paper’s tags had caught (see What Came After).

Limitations

Stated by the authors. §6 names three technical limitations. Test images of downstream benchmarks may sit in Common Crawl and inflate zero-shot results; the authors judge the risk small, citing OpenAI’s overlap analysis for CLIP and ObjectNet’s newly taken photographs, and leave measuring it, with further deduplication, to future work (see benchmark overfitting). Alt-text can be search-engine spam or keyword lists; the authors point to BASIC’s captions built from class names as one alternative. Filtering with CLIP imports CLIP’s flaws and biases, and a ViT-B/32 filter may keep more weak pairs and drop more good ones than a larger model would; the authors hypothesize that a ViT-L/14 filter would give a cleaner dataset. In the checklist they say they report no error bars because of the cost of training (item 3c).

Inconsistencies within the paper. Read closely, the paper disagrees with itself in several places:

• Table 2 gives ViT-L/14 on LAION-2B-en 75.2% on ImageNet, while §5.2.1 and Table 6 give 75.4% for the same model and data at 34 billion samples seen. The ViT-B/32 values differ the same way (Table 2: 65.7%; Table 6 at 34 billion samples: 65.50%), as do ViT-L/14 on LAION-400M (72.8% against 72.98% at 13 billion). The paper does not say which runs Table 2 reports. The ViT-L/14 LAION-2B-en model appears in Table 2 but not in Table 4’s training runs or in the per-task and retrieval Tables 7–9, and §5.2 mentions a single model trained on LAION-2B-en where Table 2 shows two. §5.2 also says 128 to 400 GPUs were used, while Table 4 lists a run on 112.
• The ViT-B/16 WIT row of Table 8 (Flickr30k) repeats its image-to-text recalls (81.9, 96.2, 98.8) as text-to-image recalls. In our reading this is a transcription error, and the text’s one exception to LAION’s retrieval advantage, text-to-image retrieval at ViT-B/16, rests on it.
• The abstract promises detection scores for watermarks, NSFW and toxic content, but the metadata columns listed in §4 and the datasheet’s list of fields (Q5) contain only the NSFW and watermark scores; the Q16 classifier appears as documentation (Figure 5), not as a released column.
• The NSFW accuracy in C.5 belongs to the EfficientNet classifier the authors say they replaced (see Method).
• The datasheet says that no preprocessing or labelling was done and, in the next sentence, that some images were removed for safety and others tagged (Q35). It rounds the subsets to 2.3, 2.2 and 1.2 billion (Q6) and the total to 5.8 billion (Q5), against 2.32, 2.26, 1.27 and 5.85 billion in the main text.
• LAIONIDE-v2’s filter drops samples whose nsfw field is NSFW or LIKELY (F.1), while §4 describes the NSFW output as a score between 0 and 1 and the datasheet speaks of boolean flags (Q16); the paper does not explain how the two relate.
• Figure 1 and §1 call LAION-400M a subset of LAION-5B, but LAION-400M had been announced on 20 August 2021 with a threshold of 0.3 (Schuhmann, 2021; Schuhmann et al., 2021), before the September 2021 to January 2022 filtering period the datasheet gives for LAION-5B (Q26). The paper does not show how the two releases relate.

Not addressed. The paper reports no measurement of duplicates, of benchmark overlap or of how many links had already died; the datasheet says only that link persistence cannot be guaranteed and that no archive of the images is kept (Q14, Q36). Its safety tags were evaluated on small hand-checked test sets, and the claim that they remove almost all unsafe results (C.7) rests on inspection of search results, not on a measurement.

What Came After

Each of the open questions was taken up by others, and the dataset itself changed; the LAION article gives the details and sources.

• Content audits. Birhane, Prabhu, Han, Boddeti and Luccioni (2023), in the NeurIPS 2023 Datasets and Benchmarks Track, compared the alt-text of samples from LAION-400M and LAION-2B-en and reported that the rate of hateful, targeted or aggressive text rose by about 12% with the larger dataset, and that filtering on the image-based NSFW score left some of that text in place.
• Duplicates. Webster, Rabin, Simon and Jurie (2023), an arXiv preprint, estimated that roughly 700 million LAION-2B-en images, about a third, are duplicates.
• Filtering as a research problem. DataComp (Gadre et al., 2023) turned the paper’s §6 question, how best to filter Common Crawl pairs, into a benchmark: participants choose a training subset from a pool of 12.8 billion Common Crawl pairs, and a fixed CLIP training run and 38 test sets score the choice.
• Withdrawal. A Stanford Internet Observatory report by David Thiel, first released on 20 December 2023, identified 3,226 LAION-5B entries as suspected child sexual abuse material (CSAM), of which 1,008 were confirmed by the Canadian Centre for Child Protection as CSAM or likely CSAM (Thiel, 2023, §1, Tables 2 and 3). It judged the collection-time filters too narrow, noting that only images both tagged NSFW and matching CLIP filters for minors had been discarded and that no image had been checked against hash lists of known material (Thiel, 2023, §5.2). LAION took its datasets down temporarily on 19 December 2023 (LAION, 2023).
• Re-LAION-5B. On 30 August 2024 LAION released Re-LAION-5B in two versions, removing 2,236 links matched against hash lists from child-protection organizations and the report’s author, among other changes (LAION, 2024). The datasheet had said that LAION-5B would not be updated, apart from removing specific samples on request (Q55). The same announcement says LAION-5B was assembled from Common Crawl data up to September 2022 and in summer 2022, dates later than both the paper’s collection period (Q26) and the release on 31 March 2022 (Q47).