REVIEW 5 major objections 5 minor 6 cited by
Meta CLIP 2: A Worldwide Scaling Recipe
T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Training CLIP on worldwide web data with properly scaled curation, model size, and seen pairs breaks the curse of multilinguality, making English and non-English data mutually beneficial.
desk verdict A strong, open recipe for multilingual CLIP that deserves a checkpoint, but the headline mutual-benefit claim needs a missing English-only 2.3x control before it is proven. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the per-language curation pipeline built on the Meta CLIP algorithm: metadata from multilingual WordNet and Wikipedia dumps in 329 languages, language identification to route each alt-text to its language's concept vocabulary, substring matching, and balancing that downsamples head concepts by a language-specific threshold t_lang. The threshold is derived under an invariance assumption that the global tail proportion p (6% of matches) measured from English data holds for every language. On the training side, the framework scales the global batch by 2.3x so the number of English pairs seen stays constant while non-English pairs are added, and uses a 900k-vocabulary multilingual tokenizer plus ViT-H/14 as the minimal viable capacity.
What would settle it
Compute the tail proportion p_lang separately for each of the 300+ languages by applying a language-specific threshold t_lang chosen from each language's own count distribution, and compare these proportions to the English-derived p. If the proportions vary substantially, the invariance assumption underpinning the recipe is false and the multilingual gains should shrink; alternatively, train a worldwide ViT-H/14 using per-language measured p instead of transferred p and check whether ImageNet and Babel-ImageNet scores change.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the trade-off between English and non-English performance in CLIP disappears when four ingredients scale together: metadata covering 300+ languages, language-specific substring matching and balancing in curation, a multilingual tokenizer and sufficient model capacity, and a training budget with seen pairs increased proportionally to the added data. Under this recipe, non-English data improves English performance and English data improves non-English performance, rather than one being sacrificed for the other. The threshold t for head/tail balancing is made language-specific by transferring the English tail proportion p to every language, and ViT-H/14 is identified as the minimal model capacity at which the curse breaks.
Load-bearing premise
The whole multilingual balancing scheme assumes the same 6% tail proportion measured in English alt-texts holds for every language, so each language's head/tail threshold is derived from English statistics rather than measured from that language's own data.
Editorial extensions
If this is right
- English-only CLIP models can be improved by adding non-English web data instead of being degraded by it, so future scaling can draw on the full web.
- Multilingual benchmarks no longer require dedicated multilingual systems with private data or machine translation; a single open recipe trained from scratch on public web data surpasses them.
- The recipe transfers to any language with a Wikipedia presence, so low-resource languages get native-language supervision and cultural coverage from their own alt-texts.
- Seen-pair scaling proportional to data growth becomes a standard training knob for worldwide contrastive models, with 2.3x as a concrete starting point.
- Models trained this way inherit broader geographic and cultural coverage, improving geo-localization and region-specific recognition.
Reading between the lines
- A direct test of the invariance assumption would compute each language's actual tail proportion from its own data; if the proportions differ, the p-transfer step is the first place to look for the remaining gap between languages.
- The 2.3x scaling factor is tied to this data's 44% English share; other data mixes should scale seen pairs by the inverse of the English fraction, and that rule could be tested across data pools.
- The same recipe should apply to text-only contrastive models or retrieval models, where the curse of multilinguality has the same structure; a test would be training a multilingual text embedder from scratch with per-language metadata and scaled pairs.
- Because the model never filters by language, cultural diversity is a byproduct of the recipe; an explicit evaluation on geographic benchmarks (Dollar Street, GeoDE, GLDv2) already shows gains, and extending XM3600-style retrieval to more languages would further check this.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes Meta CLIP 2, a recipe for training CLIP on worldwide (English + non-English) image-text pairs from scratch, with three components: per-language metadata from Wikipedia and multilingual WordNet, a language-specific curation algorithm that derives each language's head/tail threshold from a global tail proportion p computed on English, and a training framework that scales seen pairs and batch size by 2.3x while keeping English exposure approximately constant. The main empirical claim is that with ViT-H/14, worldwide training with 2.3x seen pairs outperforms the English-only 1.0x baseline on ImageNet (81.3 vs 80.4) and sets new state-of-the-art results on CVQA, Babel-ImageNet, and XM3600, thereby breaking the 'curse of multilinguality' and establishing mutual benefit between English and non-English data. The paper also includes ablations of metadata, curation, tokenizer, and cultural-diversity evaluations, and it releases code and metadata.
Significance. If validated, this is an important contribution: it is the first open, from-scratch worldwide CLIP recipe, uses no translation or distillation, and directly addresses the exhaustion of English web data. The paper's release of metadata, curation code, and models is a genuine service to the community, and the ablation structure (Tables 2 and 3) is careful and informative. However, the central causal claim about mutual benefit rests on a comparison confounded by total seen pairs and batch size, and the capacity-inflection conclusion is based on only two model sizes with no replication. These issues are load-bearing for the headline message, so the paper as written does not fully establish its central thesis.
major comments (5)
- [Table 1 and Sec. 3.4] The central mutual-benefit claim is confounded by total seen pairs and batch size. The headline comparison is H/14 Worldwide (2.3x, 29B seen pairs, batch 75,366) versus H/14 English (1.0x, 13B seen pairs, batch 32,768). These runs differ not only in data composition but also in total training signal and global batch size. The paper's rationale that 44% of 29B ~ 13B keeps English exposure constant does not control for the extra non-English signal or the larger batch. The ViT-L/14 arm (English 1.0x IN 79.5 vs Worldwide 2.3x IN 78.8) is the same confound and, as the authors note, still shows the curse; without an English-only 2.3x control run (same batch, steps, and seen pairs, achieved by re-sampling the English pool), the conclusion that non-English data helps English and that ViT-H/14 is the capacity inflection point is underdetermined, because the improvements could be a pure data-scaling or batch-size effect.
- [Sec. 4.2.1 and Table 1] All reported numbers are from single training runs with no seeds or error bars. The headline ImageNet difference is 0.8%, and several multilingual differences are a few points; without variance estimates, it is impossible to assess whether these gaps are significant relative to run-to-run noise. Please provide at least 2-3 seeds for the smaller ViT-L/14 or ViT-B/32 runs, or otherwise justify why single seeds are sufficient for the scaling claims.
- [Sec. 3.4] The 'minimal viable model capacity' claim is an extrapolation from only two model sizes, ViT-L/14 and ViT-H/14. With two points and confounded data scaling, the observation that ViT-H/14 improves while ViT-L/14 does not does not identify an 'inflection point'; it is a single comparison. An intermediate capacity (e.g., ViT-B/16 or ViT-g/14) with the same seen-pair scaling, or a study that varies capacity while holding seen pairs fixed, would be needed to support the capacity-threshold conclusion.
- [Sec. 3.3 and Table 2] The p-invariance assumption is untested. The global tail proportion p is computed from English and then transferred to all 300+ languages via p_to_t(), but the paper does not check whether p actually matches the tail proportion for each language, nor does it ablate different assumed p values. The ablation in Table 2 compares a single shared threshold with language-specific thresholds derived from the same p, so it cannot detect miscalibration of p itself. Please add a diagnostic showing the distribution of t_lang and the resulting tail proportion per language, or a sensitivity analysis around p.
- [Abstract and Sec. 4.2.1] The claim of state-of-the-art results 'without system-level confounding factors' is overstated. Although the paper greys out mSigLIP and SigLIP 2 as SoTA-aiming systems, the abstract and Section 4.2.1 still claim to surpass them with 'comparable architectures' (H/14 vs SO400M), when the comparison also involves different resolutions (224 vs 256), different training frameworks, and different loss or recipe choices. These are system-level differences, so the comparison is confounded across systems. Please reframe the comparison as cross-system rather than asserting that the results are free of system-level confounders.
minor comments (5)
- [Table 1] Column header has a typo: 'Flicker30k-200' should be 'Flickr30k-200'.
- [Algorithm 1] The comment 'balancing via independent sampling per language' is slightly misleading because the loop appends a pair at most once and breaks after the first matched entry; consider clarifying that each pair is sampled based on its first matched-entry probability.
- [Sec. 3.2] The description '40 random dates of Wikipedia snapshots' is vague; please specify the sampling procedure and why 40 snapshots were used.
- [Table 6] The 2.3x scaling is clear, but the 'Seen Pairs' row says 29B while the text says 12.8B becomes 29B; consider stating the exact multiplier (2.29x) and how the rounding was applied to avoid apparent inconsistency.
- [Appendix C] The limitation discussion in Appendix C is useful but is only in the appendix; a brief version in the main text would help calibrate the SOTA claims, since several of the multilingual benchmarks are acknowledged to have Western bias.
Circularity Check
No significant circularity: Meta CLIP 2 is an empirical scaling recipe whose claims are tested on external zero-shot benchmarks; the self-cited Meta CLIP algorithm is prior transparent work, and the p-invariance / batch-scaling concerns are unvalidated assumptions or attribution confounds, not definitional reductions.
full rationale
The paper's derivation chain is not a formal derivation but an empirical recipe. The curation threshold tlang is obtained from the stated invariance assumption (Sec. 3.3), and the 2.3x batch scaling is chosen from the measured 44% English share (Sec. 3.4); neither quantity is fitted to the reported benchmark numbers. The central claims (curse broken at ViT-H/14, mutual benefit, multilingual SoTA) are evaluated on independent zero-shot tasks such as ImageNet, Babel-ImageNet, CVQA, and XM3600, whose scores are not outputs of the curation or training equations. The reliance on Meta CLIP (Xu et al., 2024) is a transparent, code-released prior algorithm, not an unverified uniqueness theorem, and no load-bearing argument reduces to a self-citation. The closest concerns are: (1) the tail-proportion invariance p is assumed to transfer to all languages rather than validated per language, which is an empirical calibration risk, not circularity; and (2) the English 1.0x vs Worldwide 2.3x comparison also changes total seen pairs and global batch size, so the attribution of the 0.8% ImageNet gain to non-English data is underdetermined. That is an experimental-control weakness, not a definitional equivalence. The Appendix C benchmark limitation is likewise a validity caveat, not circularity. No equation in the paper reduces a predicted result to an input by construction; therefore the circularity score is 0.
Assumptions & free parameters
free parameters (3)
- English threshold t_en =
170,000 in main runs (20,000 in OpenAI CLIP setting)
- global tail proportion p =
approximately 6% tail matches for English
- global batch scaling =
2.3x (batch size 75,366 vs 32,768)
assumptions (5)
- domain assumption The same tail proportion p holds across all languages
- domain assumption Wikipedia and multilingual WordNet metadata cover visual concepts in each language
- domain assumption Language identification assigns each alt-text to the correct language-specific metadata
- domain assumption OpenAI CLIP and Meta CLIP hyperparameters transfer to worldwide data
- domain assumption Scaling seen pairs proportionally to data size is sufficient to break the curse
Cite this review
Pith. "Pith review of Meta CLIP 2: A Worldwide Scaling Recipe." pith.science (2026). https://pith.science/paper/5RQZVUR6
@misc{pith2026250722062,
author = {Pith},
title = {Pith review of: Meta CLIP 2: A Worldwide Scaling Recipe},
year = {2026},
howpublished = {\url{https://pith.science/paper/5RQZVUR6}},
note = {Machine review of arXiv:2507.22062}
}
read the original abstract
Contrastive Language-Image Pretraining (CLIP) is a popular foundation model, supporting from zero-shot classification, retrieval to encoders for multimodal large language models (MLLMs). Although CLIP is successfully trained on billion-scale image-text pairs from the English world, scaling CLIP's training further to learning from the worldwide web data is still challenging: (1) no curation method is available to handle data points from non-English world; (2) the English performance from existing multilingual CLIP is worse than its English-only counterpart, i.e., "curse of multilinguality" that is common in LLMs. Here, we present Meta CLIP 2, the first recipe training CLIP from scratch on worldwide web-scale image-text pairs. To generalize our findings, we conduct rigorous ablations with minimal changes that are necessary to address the above challenges and present a recipe enabling mutual benefits from English and non-English world data. In zero-shot ImageNet classification, Meta CLIP 2 ViT-H/14 surpasses its English-only counterpart by 0.8% and mSigLIP by 0.7%, and surprisingly sets new state-of-the-art without system-level confounding factors (e.g., translation, bespoke architecture changes) on multilingual benchmarks, such as CVQA with 57.4%, Babel-ImageNet with 50.2% and XM3600 with 64.3% on image-to-text retrieval.
Forward citations
Cited by 6 Pith papers
-
Enhancing VLM Reward Models Through Structure-Aware Fine-Tuning
SAFT adds self-supervised structure-aware losses (invariance and proportionality) via LoRA adapters to VLM rewards, improving RL policy learning and reward alignment across four control tasks.
-
Peel neighborhoods
Peel neighborhoods give a canonical, parameter-free local geometry tool in strict-negative-type finite metric spaces, enabling scalable local-dimension and singularity estimates.
-
FG-CLIP 2: A Bilingual Fine-grained Vision-Language Alignment Model
A two-stage bilingual CLIP-style model with region-text supervision and a new text-side contrastive loss outperforms prior open models on fine-grained vision-language tasks in English and Chinese.
-
Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model
Codec-guided sparse patch selection plus a lightweight speak/silent gate yields a 4B streaming VLM that is competitive on static tasks, stronger on video/spatial benchmarks, and much cheaper at inference.
-
Fine-Grained Food Image Understanding via Target-Aware Data Alignment
A data-curation pipeline—selecting target-like web images and rewriting captions with a VLM—raises food retrieval on Dishcovery Mission II to 0.653, with caption rewriting alone adding ~19%.
-
Xray-Visual Models: Scaling Vision models on Industry Scale Data
A 2-billion-parameter vision encoder trained on 15B+ image-text and billions of video-hashtag pairs reports SOTA ImageNet linear-probe, Kinetics, and retrieval numbers, but relies on proprietary data and has several v...
Reference graph
Works this paper leans on
-
[1]
Towards zero-shot cross-lingual image retrieval.arXiv preprint arXiv:2012.05107 ,
Pranav Aggarwal and Ajinkya Kale. Towards zero-shot cross-lingual image retrieval.arXiv preprint arXiv:2012.05107 ,
arXiv 2012
-
[2]
C Limitation on Benchmark High-quality benchmarks are essential for researchers to understand the efficacy of proposed changes. After decades of meticulous efforts, the community has established reliable and diverse datasets to enable research advancement in vision and multimodal areas (Deng et al., 2009; Russakovsky et al., 2015; Radford et al., 2021). H...
work page 2009
-
[9]
Learning word vectors for 157 languages
Édouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomáš Mikolov. Learning word vectors for 157 languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018),
work page 2018
-
[12]
Graph-rise: Graph-regularized image semantic embedding.arXiv preprint arXiv:1902.10814,
Da-Cheng Juan, Chun-Ta Lu, Zhen Li, Futang Peng, Aleksei Timofeev, Yi-Ting Chen, Yaxi Gao, Tom Duerig, Andrew Tomkins, and Sujith Ravi. Graph-rise: Graph-regularized image semantic embedding.arXiv preprint arXiv:1902.10814,
arXiv 1902
-
[13]
Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.International journal of computer vision , 128(7):1956–1981,
work page 1956
-
[14]
Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, and Madian Khabsa. Xlm-v: Overcoming the vocabulary bottleneck in multilingual masked language models.arXiv preprint arXiv:2301.10472,
-
[15]
12 Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie
https://openreview.net/forum?id=E18kRXTGmV. 12 Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pre-training. arXiv preprint arXiv:2112.12750 ,
-
[16]
Gabriel Oliveira dos Santos, Diego AB Moreira, Alef Iury Ferreira, Jhessica Silva, Luiz Pereira, Pedro Bueno, Thiago Sousa, Helena Maia, Nádia Da Silva, Esther Colombini, et al. Capivara: Cost-efficient approach for improving multilingual clip performance on low-resource languages.arXiv preprint arXiv:2310.13683 ,
Show all 29 references
-
[17]
Laion-400m: Open dataset of clip-filtered 400 million image-text pairs
Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 ,
-
[18]
Laion-5b: An open large-scale dataset for training next generation image-text models
Christoph Schuhmann, Romain Beaumont, Cade W Gordon, Ross Wightman, Theo Coombes, Aarush Katta, Clayton Mullis, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022a. Chr...
-
[20]
Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,
Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,
-
[21]
Gemma: Open models based on gemini research and technology
Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295 ,
-
[22]
Crossmodal-3600: A massively multilingual multimodal evaluation dataset
13 Ashish V Thapliyal, Jordi Pont Tuset, Xi Chen, and Radu Soricut. Crossmodal-3600: A massively multilingual multimodal evaluation dataset. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 715–729,
2022
-
[24]
Will we run out of data? limits of llm scaling based on human-generated data.arXiv preprint arXiv:2211.04325 ,
Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn. Will we run out of data? limits of llm scaling based on human-generated data.arXiv preprint arXiv:2211.04325 ,
-
[25]
Nllb-clip–train performant multilingual image retrieval model on a budget.arXiv preprint arXiv:2309.01859,
Alexander Visheratin. Nllb-clip–train performant multilingual image retrieval model on a budget.arXiv preprint arXiv:2309.01859,
-
[26]
Scaling pre-training to one hundred billion data for vision language models.arXiv preprint arXiv:2502.07617 ,
Xiao Wang, Ibrahim Alabdulmohsin, Daniel Salz, Zhe Li, Keran Rong, and Xiaohua Zhai. Scaling pre-training to one hundred billion data for vision language models.arXiv preprint arXiv:2502.07617 ,
-
[27]
Hu Xu, Saining Xie, Xiaoqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer
Accessed: 2025-05-15. Hu Xu, Saining Xie, Xiaoqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying CLIP data. InThe Twelfth International Conference on Learning Representations,
2025
-
[28]
Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel
https://openreview.net/forum?id=5BCFlnfE1g. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mt5: A massively multilingual pre-trained text-to-text transformer.arXiv preprint arXiv:2010.11934 ,
2010 arXiv
-
[2009]
Scaling language-free visual representation learning.arXiv preprint arXiv:2504.01017 ,
David Fan, Shengbang Tong, Jiachen Zhu, Koustuv Sinha, Zhuang Liu, Xinlei Chen, Michael Rabbat, Nicolas Ballas, Yann LeCun, Amir Bar, et al. Scaling language-free visual representation learning.arXiv preprint arXiv:2504.01017 ,
-
[2015]
Unsupervised cross-lingual representation learning at scale
Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 ,
1911 arXiv
-
[2017]
Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818 ,
Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818 ,
-
[2018]
Distilling the knowledge in a neural network, 2015.https://arxiv
Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015.https://arxiv. org/abs/1503.02531. Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, ...
2015 arXiv
-
[2019]
Imagenet: A large-scale hierarchical image database
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pages 248–255. Ieee,
2009
-
[2020]
PaLI: A jointly-scaled multilingual language-image model
Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James...
-
[2021]
If you use this software, please cite it as below
https://doi.org/10.5281/zenodo.5143773. If you use this software, please cite it as below. Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning wit...
-
[2022]
Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786 ,
Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and ...
-
[2023]
Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi, Jathushan Rajasegaran, Hanoona Rasheed, et al
https://arxiv.org/abs/2308.12966. Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi, Jathushan Rajasegaran, Hanoona Rasheed, et al. Perception encoder: The best visual embeddings are not at the output of the network. arXiv pre...
-
[2024]
The llama 3 herd of models
Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,
-
[2025]
Data filtering networks
Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425 ,
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.