Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

FlexOlmo: Open Language Models for Flexible Data Use

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FlexOlmo shows that language-model experts trained on separate closed datasets can be merged into one router-guided mixture with no joint training, improving over the public-only model by 41% on average.

desk verdict A genuinely new expert-merging recipe with solid results; the overclaimed opt-out guarantee and the unexplained 41% number need referee attention. read the letter →

arxiv 2507.07024 v4 pith:K635JD2H submitted 2025-07-09 cs.CL cs.AI

Weijia Shi , Akshita Bhagia , Kevin Farhat , Niklas Muennighoff , Pete Walsh , Jacob Morrison , Dustin Schwenk , Shayne Longpre
show 15 more authors
Jake Poznanski Allyson Ettinger Daogao Liu Margaret Li Dirk Groeneveld Mike Lewis Wen-tau Yih Luca Soldaini Kyle Lo Noah A. Smith Luke Zettlemoyer Pang Wei Koh Hannaneh Hajishirzi Ali Farhadi Sewon Min
This is my paper
classification cs.CLcs.AI
keywords mixture-of-expertsmodelmergingrouterlearningdataprivacycloseddatasetsopt-outdistributedtraininglanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to show that a language model can be built from parts trained by different data owners without any of them sharing their data. Each owner trains a small expert network and a per-expert router embedding on their own closed dataset, using a shared public model as a frozen anchor; the pieces are then combined into a mixture-of-experts model whose router is assembled from those embeddings. The central claim is that this no-joint-training composition works: on 31 tasks it improves over the public-only model by 41% relative, and it lets users add or remove any expert at inference, so a dataset can be opted out by removing its module. If correct, this gives regulated organizations a way to contribute sensitive or licensed data to shared models while keeping the data local and controlling when it is used.

What carries the argument

The load-bearing object is the domain-informed router built from per-expert router embeddings. Each row is initialized as the average embedding of a small sample of the owner's dataset, $r_i = \frac{1}{|S_i|}\sum_{d_k \in S_i} E(d_k)$, produced by an off-the-shelf text embedder, then finetuned during pairwise coordinated training alongside the expert FFN while the public row $r_{\mathrm{pub}}$ stays frozen. Stacking these rows forms the full router matrix $W_r$, and the negative bias $b_i$ makes each expert's decision boundary tighter, so the composed rule $r_i \cdot x + b_i > r_{\mathrm{pub}} \cdot x$ approximates a multiclass choice among all experts. This router is what lets independent modules coordinate with no joint training and what makes opt-out a structural operation: remove the row, remove the expert.

What would settle it

Train FlexOlmo-style experts on a dozen disjoint, well-separated domains and also train a standard end-to-end MoE router on the union of those domains. If the composed pairwise router disagrees with the joint router on most in-domain tokens, or if its average downstream score is no better than always selecting the public expert, the central router-composition claim fails. A cheaper check uses triples: if the learned pairwise preference is not transitive (i.e., expert A beats public, expert B beats public, but B beats A while the composed argmax still routes to A), the merged router is silently inconsistent.

Watch

Extended reading notes

Core claim

The central discovery is that independent expert modules can be made mergeable without joint training by training each one against the same frozen public model and giving each expert its own router embedding, learned only from that expert's data. At merge time the router matrix is formed by stacking the per-expert rows, and each specialized expert is selected when $r_i \cdot x + b_i > r_{\mathrm{pub}} \cdot x$, a negative bias tightening the 'expert vs. public' boundary into a more selective 'expert vs. everything else' boundary. Because the shared attention layers and the public router row are frozen, all binary classifiers live in one coordinate system, and because each expert owns only its FFN and router row, removing a module removes the parameters that saw that dataset. The paper reports that the resulting eight-expert model with 20 billion active parameters improves over the public model by 41% relative, beats the strongest prior merging baseline by 10.1%, and roughly matches an unrestricted MoE trained on all data at equal FLOPs.

Load-bearing premise

Everything depends on the assumption that many one-vs-public routing decisions, made by classifiers that never saw each other's data, can be stitched into one many-way router that behaves like a router trained on all the data together.

Editorial extensions

If this is right

  • A data owner can join the model asynchronously: train one expert against the public anchor, then add its router row; no retraining of existing experts or central pooling is needed.
  • Opt-out is structural: deleting an expert module and its router row takes that dataset's trained parameters out of the model, which the paper states as a 'complete removal' guarantee.
  • Prior merging families—weight averaging (model soup) and output ensembling (BTM)—are expected to underperform because they lack per-layer learned routing; FlexOlmo beats BTM by 10.1% on average.
  • FlexOlmo can extend an already strong pretrained model: a three-expert version built on a 4-trillion-token checkpoint beats its dense counterpart at matched training FLOPs, indicating specialization without catastrophic forgetting.
  • The recipe is an alternative to federated learning for language modeling, avoiding synchronized communication and joint data access while allowing each owner to apply differential privacy locally if desired.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper's results, the pairwise-vs-multiclass router gap is the natural scaling bottleneck: if the negative-bias composition degrades as expert count grows, one would expect gains to flatten or invert beyond the eight experts tested, and a study with dozens of domains on a few tasks could test that.
  • The 'complete removal' guarantee is architectural rather than a formal privacy statement; a targeted membership-inference test on the shared frozen attention layers would show how much, if anything, leaks through the anchor itself.
  • The router-embedding mechanism suggests a general recipe for modular continual learning: new skills can be bolted on by any party that can see the public anchor, which may be useful for incremental scientific collaboration across institutions.
  • Because differential privacy is orthogonal to the architecture, each contributor can choose a different privacy budget; comparing the utility of FlexOlmo with DP-trained experts versus no DP is an obvious next experiment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces FlexOlmo, a mixture-of-experts architecture in which each expert is trained independently on a closed dataset using a frozen public anchor, and experts are later combined through a router built from per-expert embeddings. The authors claim that this enables distributed training without data sharing, flexible data opt-in/opt-out at inference, and no joint training. They curate FlexMix, train models up to 37B total parameters, evaluate on 31 tasks, and report large gains over the public model and over prior merging baselines, along with ablations, routing analyses, a data-extraction study, and a scaling experiment.

Significance. If the empirical claims hold after the corrections below, this is a valuable contribution to training language models under data-sharing restrictions. The paper's strengths include releasing the model and code, evaluating on a broad set of 31 tasks with controlled comparisons, isolating the effect of each design component through ablations, honestly measuring data-extraction risk, and testing the recipe at a larger scale. The main weaknesses are that two headline claims—the 41% relative improvement and the 'complete removal' guarantee—are not supported by the paper's own tables and training procedure, and the central router-composition assumption is acknowledged but not directly validated.

major comments (4)
  1. [§3.1 vs. §3.3.3 and Appendix A.2] The claim that removing any module Mi from Mfinal guarantees complete removal of its associated data Di is contradicted by the optional router training used in the main results. The proxy set D̂i is selected with a binary classifier trained on Di versus Dpub, and then rpub is tuned jointly with all ri on the union of these proxy sets and Dpub. Because D̂i is a function of Di, the retained rpub carries Di-dependent information even after Mi is deleted. Table 2's 'FLEX OLMO' row uses this optional training, so the guarantee as stated does not hold for the reported final model. Please either scope the guarantee to the configuration without optional router training or modify the procedure so that rpub is not influenced by Di.
  2. [Abstract and §5.1] The reported 'average 41% relative improvement' over the public model is not supported by Tables 1 and 2. Using the average scores in those tables, the relative gains are (47.8−36.9)/36.9 ≈ 29.5% and (52.4−42.4)/42.4 ≈ 23.6%, respectively. Please state exactly how the 41% figure is computed (for example, as a mean of per-task relative improvements) or correct the number in the abstract and §5.1.
  3. [Abstract and §3.3.3] The statement that FlexOlmo integrates experts 'without any joint training' is at odds with the optional router tuning step, which jointly tunes r1,...,rn and rpub on the combined proxy sets and Dpub. If the claim is intended to mean only that there is no joint training on the closed datasets themselves, that qualification must be stated explicitly; as written, the final system does include a joint training stage.
  4. [Appendix D] The core assumption that pairwise-trained binary routers can be composed into a near-optimal multiclass router is acknowledged in Appendix D as an open question and is supported only by an intuitive argument. Because this assumption underlies the 'no joint training' contribution and the merging gains, please add a controlled experiment that isolates router quality: train the same experts, then compare the composed router against a router trained jointly on the combined proxy/public data under otherwise identical conditions. The unrestricted MoE baseline confounds router quality with expert training and data allocation, so it does not resolve this question.
minor comments (4)
  1. [§4.4] The token accounting is ambiguous: 'continued-pretraining for 50 billion tokens on their own data (totaling 400B tokens across all experts)' does not match the seven closed sets (7 × 50B = 350B) unless the public expert is also counted; please clarify whether the public expert receives an additional 50B tokens.
  2. [§5.4] There are two typos in this section: 'againsts' should be 'against', and 'starting from a a checkpoint' should be 'starting from a checkpoint'.
  3. [Appendix A.2] When reporting that the proxy-selection classifier 'quickly achieved an accuracy above 95% across all datasets considered', please specify the evaluation split (held-out data) and report per-dataset accuracies, since the proxy selection quality is directly relevant to the guarantee discussed in §3.1.
  4. [§3.3.3 and §3.1] The problem setup in §3.1 grants each data owner access only to Mpub, but the optional router training in §3.3.3 requires access to Dpub to construct the proxy sets; please clarify this assumption, as it affects the practical feasibility of the final reported configuration.

Circularity Check

1 steps flagged · score 6.0 of 10

FlexOlmo's strict opt-out guarantee is falsified by its own optional router tuning: proxy sets are selected with classifiers trained on Di and then used to tune rpub, so Di influence survives Mi removal.

  1. fitted input called prediction [Section 3.1; Section 3.3.3; Appendix A.2; Table 2]
    "removing any module Mi from Mfinal guarantees complete removal of its associated data Di. ... To construct ˆDi, we train a binary classifier to distinguish Di from Dpub and select public samples with the highest predicted likelihood of belonging to Di. ... The classifier is trained on a balanced dataset of 500,000 samples (250,000 documents from each source - public and private). After merging, we tune the router embeddings r1, · · ·, rn, rpub on the combined set ˆD1, · · ·, ˆDn, and Dpub, sampled uniformly."

    The 'complete removal' guarantee is asserted as a structural property of deleting Mi, but the reported final model uses the optional router-tuning step. The paper's own chain is Di → binary classifier trained on Di vs Dpub → proxy set D̂i → joint tuning of rpub and all ri on {D̂i} ∪ Dpub. Thus rpub is a fitted function of Di. Deleting Mi removes the expert FFN and ri, but rpub remains and carries Di-derived routing information, so Di influence persists in the retained model. The guarantee is therefore not derived from the architecture; it is a fitted-parameter dependence presented as an opt-out guarantee. The only variant for which the guarantee can hold is 'FLEX OLMO (no RT)', which is not the reported final model in Table 2.

full rationale

The core empirical content of the paper is not circular: router embeddings are initialized from an external embedder (GRIT) and fine-tuned per expert on each Di; the reported gains are measured against external or reimplemented baselines; and the authors' own prior artifacts (OLMo 2, GRIT, Dolma) are used as components rather than as evidence for the central claim. The one structural circularity is the opt-out guarantee. Section 3.1 states it as a formal requirement, but Section 3.3.3 and Appendix A.2 show that the optionally tuned router rpub is a fitted function of every Di: Di-trained proxy classifiers select D̂i, and rpub is jointly tuned on the combined proxy sets. Because the reported final model uses this optional tuning, removing Mi leaves an Di-dependent parameter in place, so the 'complete removal' guarantee is not a consequence of the architecture—it is a relabeling of a fitted parameter as a structural guarantee, and the paper's own procedure falsifies it. The same optional step also makes the abstract's literal claim of integration 'without any joint training' inaccurate for the final model, since rpub and all ri are tuned on the combined proxy sets; that is a consistency issue rather than a separate derivation circle. Appendix D's explicit open question about composing binary classifiers into a multiclass router is an honest limitation, not a hidden circularity. Overall, the central empirical contribution retains independent content, but one load-bearing guarantee reduces by construction to a fitted-parameter dependence, giving a partial-circularity score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The empirical results depend on several domain assumptions: dataset disjointness, the compositionality of pairwise routers, the utility of GRIT embeddings, and the representativeness of proxy data. There are no fitted physical constants or invented entities.

free parameters (4)
  • Per-expert training tokens = 50B tokens
    Each expert is continued-pretrained for 50B tokens on its closed dataset; this budget is chosen by hand and affects specialization vs. forgetting.
  • Number of active experts at inference = 4 of 8
    Figure 3 shows performance plateaus at 4 active experts; the final model uses 4, which trades inference cost for accuracy.
  • Router embedding sample size = 1000 documents per dataset
    Domain embeddings are initialized by averaging GRIT embeddings over 1000 sampled documents per source; the sample size is chosen by hand.
  • Proxy set size limit = less than 0.01 times |Di|
    Optional router tuning uses a proxy subset of public data capped at 1% of the closed set size; the threshold is a design choice.
assumptions (5)
  • domain assumption The closed datasets Di are disjoint from the public dataset Dpub and from each other.
    Stated in Section 4.1 ('designed to be disjoint from each other and from the Public Mix'). The opt-out guarantee and the attribution of gains to specific experts depend on this non-overlap.
  • ad hoc to paper Pairwise-trained binary routers can be composed into a near-optimal multiclass router.
    Appendix D frames this as an open question and provides only an intuitive argument based on a shared frozen anchor and a negative bias. The central result relies on this composition working.
  • domain assumption The off-the-shelf GRIT embedder provides domain embeddings that are a good initialization for router embeddings.
    Used in Section 3.3.2 and Table 3; the choice of embedder affects routing quality, and no theoretical guarantee is given.
  • domain assumption Proxy subsets of public data can approximate the distribution of closed datasets well enough to improve router tuning.
    Assumed in Section 3.3.3; the proxy selection uses a binary classifier trained on Di, but the representativeness is only validated empirically on benchmarks.
  • domain assumption The frozen public attention layers provide a stable shared coordinate system for independently trained experts.
    The training algorithm in Section 3.3.1 relies on freezing attention and the public expert so that expert FFNs learn compatible representations. This is a design assumption, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlexOlmo: Open Language Models for Flexible Data Use." pith.science (2026). https://pith.science/paper/K635JD2H

@misc{pith2026250707024,
  author       = {Pith},
  title        = {Pith review of: FlexOlmo: Open Language Models for Flexible Data Use},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K635JD2H}},
  note         = {Machine review of arXiv:2507.07024}
}
read the original abstract

We introduce FlexOlmo, a new class of language models (LMs) that supports (1) distributed training without data sharing, where different model parameters are independently trained on closed datasets, and (2) data-flexible inference, where these parameters along with their associated data can be flexibly included or excluded from model inferences with no further training. FlexOlmo employs a mixture-of-experts (MoE) architecture where each expert is trained independently on closed datasets and later integrated through a new domain-informed routing without any joint training. FlexOlmo is trained on FlexMix, a corpus we curate comprising publicly available datasets alongside seven domain-specific sets, representing realistic approximations of closed sets. We evaluate models with up to 37 billion parameters (20 billion active) on 31 diverse downstream tasks. We show that a general expert trained on public data can be effectively combined with independently trained experts from other data owners, leading to an average 41% relative improvement while allowing users to opt out of certain data based on data licensing or permission requirements. Our approach also outperforms prior model merging methods by 10.1% on average and surpasses the standard MoE trained without data restrictions using the same training FLOPs. Altogether, this research presents a solution for both data owners and researchers in regulated industries with sensitive or protected data. FlexOlmo enables benefiting from closed data while respecting data owners' preferences by keeping their data local and supporting fine-grained control of data access during inference.

Figures

Figures reproduced from arXiv: 2507.07024 by the authors.

Figure 1
Figure 1. An overview of FLEXOLMO. Data owners can contribute without sharing the data by training their own expert modules (FFNs and router embeddings) with a shared public model as an anchor point. At inference, these modules are integrated into a MoE model via a novel router embedding concatenation. This design enables flexible inclusion or exclusion of experts and strict opt-out guarantees, e.g., Github data can be exclud… view at source ↗
Figure 2
Figure 2. Routing pattern analysis. We visualize how text from different domains activate experts (four experts activated). The horizontal gray lines indicate uniform routing. only. However, their performance is inconsistent: model soup and BTX are generally weak,3 while prompt-based routing is highly unstable: it performs well when the classifier selects the correct expert, but degrades sharply when it does not. Among the ba… view at source ↗
Figure 3
Figure 3. Effect of active expert count on MMLU performance. Model performance stabilizes after activating four experts [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Statistics of our data mix (descriptions in §4.1). Name # Tokens (B) Public Mix 2.37 × 103 News 158.0 Creative Writing 201.9 Math 20.3 StarCoder 83.0 Academic 58.6 Educational text 102.2 Reddit 9.9 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: During training, the negative bias shifts the decision boundary. So that a more selective subset of data will be used to train the expert corresponding to C1. To ensure each classifier fi learns a better decision boundary hi , we propose two key techniques: Freezing rp…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. User Privacy and Large Language Models: An Analysis of Frontier Developers' Privacy Policies

    cs.CY 2025-09 conditional novelty 6.0 of 10

    All six leading U.S. AI chatbot developers, as of May 2025, appear to train their models on users' chat data by default, often without clear opt-out options.

Reference graph

Works this paper leans on

127 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    TOFU: A task of fictitious unlearning for LLMs

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary Chase Lipton, and J Zico Kolter. TOFU: A task of fictitious unlearning for LLMs. In First Conference on Language Modeling, 2024

    2024

  2. [2]

    Muse: Machine unlearning six- way evaluation for language models

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A Smith, and Chiyuan Zhang. Muse: Machine unlearning six- way evaluation for language models. arXiv preprint arXiv:2407.06460, 2024

    arXiv 2024

  3. [3]

    Machine unlearning doesn’t do what you think: Lessons for generative ai policy, research, and practice

    A Feder Cooper, Christopher A Choquette-Choo, Miranda Bogen, Matthew Jagielski, Katja Filippova, Ken Ziyu Liu, Alexandra Chouldechova, Jamie Hayes, Yangsibo Huang, Niloofar Mireshghallah, et al. Machine unlearning doesn’t do what you think: Lessons for generative ai policy, research, and practice. arXiv preprint arXiv:2412.06966, 2024

    arXiv 2024

  4. [4]

    Federated learning: Strategies for improving communication efficiency

    Jakub Konecn`y, H Brendan McMahan, X Yu Felix, Peter Richtárik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. CoRR, 2016

    2016

  5. [5]

    Communication-Efficient Learning of Deep Networks from Decentralized Data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54 of Proceedings of Machine Learning Research, pages 1273...

    2017

  6. [6]

    Kang Wei, Jun Li, Ming Ding, Chuan Ma, Howard Hua Yang, Farokhi Farhad, Shi Jin, Tony Q. S. Quek, and H. Vincent Poor. Federated learning with differential privacy: Algorithms and performance analysis. IEEE Transactions on Information Forensics and Security, 15:3454– 3469, 2019

    2019

  7. [7]

    Trading off privacy, utility, and efficiency in federated learning

    Xiaojin Zhang, Yan Kang, Kai Chen, Lixin Fan, and Qiang Yang. Trading off privacy, utility, and efficiency in federated learning. ACM Transactions on Intelligent Systems and Technology, 14:1 – 32, 2022

    2022

  8. [8]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    arXiv 2017

Show all 127 references
  1. [9]

    Olmoe: Open mixture-of- experts language models

    Niklas Muennighoff, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Jacob Morrison, Sewon Min, Weijia Shi, Pete Walsh, Oyvind Tafjord, Nathan Lambert, et al. Olmoe: Open mixture-of- experts language models. arXiv preprint arXiv:2409.02060, 2024

    2024 arXiv

  2. [10]

    A survey on model moerging: Recycling and routing among specialized experts for collaborative learning

    Prateek Yadav, Colin Raffel, Mohammed Muqeeth, Lucas Caccia, Haokun Liu, Tianlong Chen, Mohit Bansal, Leshem Choshen, and Alessandro Sordoni. A survey on model moerging: Recycling and routing among specialized experts for collaborative learning. CoRR, 2024

    2024

  3. [11]

    Smith, and Luke Zettlemoyer

    Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A. Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models, 2022

    2022

  4. [12]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo- Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inf...

    2022

  5. [13]

    Generative representational instruction tuning, 2024

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning, 2024. 12

    2024

  6. [14]

    Project aurora: the power of data, technology and collaboration to combat money laundering across institutions and borders

    BIS Innovation Hub Nordic Centre. Project aurora: the power of data, technology and collaboration to combat money laundering across institutions and borders. Technical report. Occasional publication No. 66

  7. [15]

    A survey on federated learning

    Chen Zhang, Yu Xie, Hang Bai, Bin Yu, Weihong Li, and Yuan Gao. A survey on federated learning. Knowledge-Based Systems, 216:106775, 2021

    2021

  8. [16]

    Data sovereignty: A review

    Patrik Hummel, Matthias Braun, Max Tretter, and Peter Dabrock. Data sovereignty: A review. Big Data & Society, 8(1):2053951720982012, 2021

    2021

  9. [17]

    Databricks lakeguard: Supporting fine-grained access control and multi-user capabilities for apache spark workloads

    Martin Grund, Stefania Leone, Herman van Hövell, Sven Wagner-Boysen, Sebastian Hillig, Hyukjin Kwon, David Lewis, Jakob Mund, Polo-Francois Poli, Lionel Montrieux, et al. Databricks lakeguard: Supporting fine-grained access control and multi-user capabilities for apache spark ...

    2025

  10. [18]

    Productivity assessment of neural code completion

    Albert Ziegler, Eirini Kalliamvakou, X Alice Li, Andrew Rice, Devon Rifkin, Shawn Simis- ter, Ganesh Sittampalam, and Edward Aftandilian. Productivity assessment of neural code completion. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, pa...

    2022

  11. [19]

    Foundation models and fair use

    Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and fair use. Journal of Machine Learning Research, 24(400):1–79, 2023

    2023

  12. [20]

    Consent in crisis: The rapid decline of the ai data commons

    Shayne Longpre, Robert Mahari, Ariel Lee, Campbell Lund, Hamidah Oderinwale, William Brannon, Nayan Saxena, Naana Obeng-Marnu, Tobin South, Cole Hunter, et al. Consent in crisis: The rapid decline of the ai data commons. Advances in Neural Information Processing Systems, 37:10...

    2024

  13. [21]

    Art and the science of generative ai

    Ziv Epstein, Aaron Hertzmann, Investigators of Human Creativity, Memo Akten, Hany Farid, Jessica Fjeld, Morgan R Frank, Matthew Groh, Laura Herman, Neil Leach, et al. Art and the science of generative ai. Science, 380(6650):1110–1111, 2023

    2023

  14. [22]

    The foundation model transparency index

    Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, and Percy Liang. The foundation model transparency index. arXiv preprint arXiv:2310.12941, 2023

    2023 arXiv

  15. [23]

    Foundation models and copyright questions, 2023

    Peter Henderson, Xuechen Li, Dan Jurafsky, Tatsunori Hashimoto, Mark A Lemley, and Percy Liang. Foundation models and copyright questions, 2023

    2023

  16. [24]

    Bollacker, Tongshuang Wu, Luis Villa, Sandy Pentland, and Sara Hooker

    Shayne Longpre, Robert Mahari, Anthony Chen, Naana Obeng-Marnu, Damien Sileo, William Brannon, Niklas Muennighoff, Nathan Khazam, Jad Kabbara, Kartik Perisetla, Xinyi Wu, Enrico Shippole, Kurt D. Bollacker, Tongshuang Wu, Luis Villa, Sandy Pentland, and Sara Hooker. A large-sc...

    2024

  17. [25]

    On the opportunities and risks of foundation models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  18. [26]

    The data provenance initiative: A large scale audit of dataset licensing & attribution in ai.arXiv preprint arXiv:2310.16787, 2023

    Shayne Longpre, Robert Mahari, Anthony Chen, Naana Obeng-Marnu, Damien Sileo, William Brannon, Niklas Muennighoff, Nathan Khazam, Jad Kabbara, Kartik Perisetla, et al. The data provenance initiative: A large scale audit of dataset licensing & attribution in ai.arXiv preprint a...

    arXiv

  19. [27]

    Into the laion’s den: Investigating hate in multimodal datasets

    Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion’s den: Investigating hate in multimodal datasets. Advances in neural information processing systems, 36:21268–21284, 2023

    2023

  20. [28]

    Documenting large webtext corpora: A case study on the colossal clean crawled corpus

    Jesse Dodge, Maarten Sap, Ana Marasovi´c, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. Documenting large webtext corpora: A case study on the colossal clean crawled corpus. In Proceedings of the 2021 Conference on Empirical Methods in N...

    2021

  21. [29]

    Detecting pretraining data from large language models

    Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. Detecting pretraining data from large language models. In The Twelfth International Conference on Learning Representations

  22. [30]

    Advances and open problems in federated learning

    Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Ar- jun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(...

    2021

  23. [31]

    Federated optimization in heterogeneous networks

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429–450, 2020

    2020

  24. [32]

    Fedlora: When personalized federated learning meets low-rank adaptation

    Xinghao Wu, Xuefeng Liu, Jianwei Niu, Haolin Wang, Shaojie Tang, and Guogang Zhu. Fedlora: When personalized federated learning meets low-rank adaptation. 2024

    2024

  25. [33]

    Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning

    Haodong Zhao, Wei Du, Fangqi Li, Peixuan Li, and Gongshen Liu. Fedprompt: Communication-efficient and privacy-preserving prompt tuning in federated learning. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 1–5. IEEE, 2023

    2023

  26. [34]

    Federated lora with sparse communication

    Kevin Kuo, Arian Raje, Kousik Rajesh, and Virginia Smith. Federated lora with sparse communication. arXiv preprint arXiv:2406.05233, 2024

    2024 arXiv

  27. [35]

    Public-key cryptosystems based on composite degree residuosity classes

    Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In Proceedings of the 17th International Conference on Theory and Application of Cryptographic Techniques, EUROCRYPT’99, page 223–238, Berlin, Heidelberg, 1999. Springer-Verlag

    1999

  28. [36]

    The algorithmic foundations of differential privacy

    Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends® in Theoretical Computer Science, 9(3–4):211–407, 2014

    2014

  29. [37]

    Attack of the tails: Yes, you really can backdoor federated learning

    Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Li...

    2020

  30. [38]

    Exploring adversarial attacks in federated learning for medical imaging

    Erfan Darzi, Florian Dubost, Nanna M Sijtsema, and Peter MA van Ooijen. Exploring adversarial attacks in federated learning for medical imaging. IEEE Transactions on Industrial Informatics, 2024

    2024

  31. [39]

    Federated learning with buffered asynchronous aggregation

    John Nguyen, Kshitiz Malik, Hongyuan Zhan, Ashkan Yousefpour, Mike Rabbat, Mani Malek, and Dzmitry Huba. Federated learning with buffered asynchronous aggregation. In International conference on artificial intelligence and statistics, pages 3581–3607. PMLR, 2022

    2022

  32. [40]

    Building machine learning models like open source software

    Colin Raffel. Building machine learning models like open source software. Communications of the ACM, 66(2):38–40, 2023

    2023

  33. [41]

    Heterogeneous swarms: Jointly optimizing model roles and weights for multi-llm systems, 2025

    Shangbin Feng, Zifeng Wang, Palash Goyal, Yike Wang, Weijia Shi, Huang Xia, Hamid Palangi, Luke Zettlemoyer, Yulia Tsvetkov, Chen-Yu Lee, and Tomas Pfister. Heterogeneous swarms: Jointly optimizing model roles and weights for multi-llm systems, 2025

    2025

  34. [42]

    Adaptersoup: Weight averaging to improve generalization of pretrained language models

    Alexandra Chronopoulou, Matthew E Peters, Alexander Fraser, and Jesse Dodge. Adaptersoup: Weight averaging to improve generalization of pretrained language models. arXiv preprint arXiv:2302.07027, 2023

    2023 arXiv

  35. [43]

    Collective model intelligence requires compatible specialization

    Jyothish Pari, Samy Jelassi, and Pulkit Agrawal. Collective model intelligence requires compatible specialization. ArXiv, abs/2411.02207, 2024

    2024 arXiv

  36. [44]

    Model swarms: Collaborative search to adapt llm experts via swarm intelligence

    Shangbin Feng, Zifeng Wang, Yike Wang, Sayna Ebrahimi, Hamid Palangi, Lesly Miculicich, Achin Kulshrestha, Nathalie Rauschmayr, Yejin Choi, Yulia Tsvetkov, et al. Model swarms: Collaborative search to adapt llm experts via swarm intelligence. arXiv e-prints , pages arXiv–2410,...

    2024

  37. [45]

    Ties- merging: Resolving interference when merging models

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties- merging: Resolving interference when merging models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36,...

    2023

  38. [46]

    Merging models with fisher-weighted averaging

    Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 17703–17716. Curran Associates, Inc., 2022

    2022

  39. [47]

    Editing models with task arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Ha- jishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023

    2023

  40. [48]

    Evolutionary optimization of model merging recipes

    Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024

    2024 arXiv

  41. [49]

    Adamerging: Adaptive model merging for multi-task learning

    Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth Interna- tional Conference on Learning Representations, 2024

    2024

  42. [50]

    Smith, Hannaneh Hajishirzi, Pang Wei Koh, Jesse Dodge, and Pradeep Dasigi

    Jacob Morrison, Noah A. Smith, Hannaneh Hajishirzi, Pang Wei Koh, Jesse Dodge, and Pradeep Dasigi. Merge to learn: Efficiently adding skills to language models with model merging, 2024

    2024

  43. [51]

    When one llm drools, multi-llm collaboration rules, 2025

    Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, and Yulia Tsvetkov. When one llm drools, multi-llm collaboration rules, 2025

    2025

  44. [52]

    Branch-train-merge: Embarrassingly parallel training of expert language models

    Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022

    2022 arXiv

  45. [53]

    Scaling expert language models with unsupervised domain discovery.arXiv preprint arXiv:2303.14177, 2023

    Suchin Gururangan, Margaret Li, Mike Lewis, Weijia Shi, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Scaling expert language models with unsupervised domain discovery.arXiv preprint arXiv:2303.14177, 2023

    2023 arXiv

  46. [54]

    Demix layers: Disentangling domains for modular language modeling

    Suchin Gururangan, Mike Lewis, Ari Holtzman, Noah A Smith, and Luke Zettlemoyer. Demix layers: Disentangling domains for modular language modeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2022

  47. [55]

    Branch-train-mix: Mixing expert llms into a mixture-of-experts llm

    Sainbayar Sukhbaatar, Olga Golovneva, Vasu Sharma, Hu Xu, Xi Victoria Lin, Baptiste Rozière, Jacob Kahn, Daniel Li, Wen-tau Yih, Jason Weston, et al. Branch-train-mix: Mixing expert llms into a mixture-of-experts llm. arXiv preprint arXiv:2403.07816, 2024

    2024 arXiv

  48. [56]

    Scalable multi-domain adaptation of language models using modular experts

    Peter Schafhalter, Shun Liao, Yanqi Zhou, Chih-Kuan Yeh, Arun Kandoor, and James Laudon. Scalable multi-domain adaptation of language models using modular experts. arXiv preprint arXiv:2410.10181, 2024

    2024 arXiv

  49. [57]

    Bts: Harmonizing spe- cialized experts into a generalist llm

    Qizhen Zhang, Prajjwal Bhargava, Chloe Bi, Chris X Cai, Jakob Foerster, Jeremy Fu, Punit Singh Koura, Ruan Silva, Sheng Shen, Emily Dinan, et al. Bts: Harmonizing spe- cialized experts into a generalist llm. arXiv preprint arXiv:2502.00075, 2025

    2025 arXiv

  50. [58]

    Ustun, and Acyr F

    Qizhen Zhang, Nikolas Gritsch, Dwaraknath Gnaneshwar, Simon Guo, David Cairuz, Bharat Venkitesh, Jakob Foerster, Phil Blunsom, Sebastian Ruder, A. Ustun, and Acyr F. Locatelli. Bam! just like that: Simple and efficient parameter upcycling for mixture of experts. ArXiv, abs/240...

    2024 arXiv

  51. [59]

    Mod- uleformer: Modularity emerges from mixture-of-experts

    Yikang Shen, Zheyu Zhang, Tianyou Cao, Shawn Tan, Zhenfang Chen, and Chuang Gan. Mod- uleformer: Modularity emerges from mixture-of-experts. arXiv preprint arXiv:2306.04640, 2023. 15

    2023 arXiv

  52. [60]

    Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning

    Ted Zadouri, Ahmet Üstün, Arash Ahmadian, Beyza Ermi¸ s, Acyr Locatelli, and Sara Hooker. Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning. arXiv preprint arXiv:2309.05444, 2023

    2023 arXiv

  53. [61]

    Sira: Sparse mixture of low rank adaptation, 2023

    Yun Zhu, Nevan Wichers, Chu-Cheng Lin, Xinyi Wang, Tianlong Chen, Lei Shu, Han Lu, Canoee Liu, Liangchen Luo, Jindong Chen, and Lei Meng. Sira: Sparse mixture of low rank adaptation, 2023

    2023

  54. [62]

    LoRAMoE: Alleviating world knowledge forgetting in large language models via MoE-style plugin

    Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Wei Shen, Limao Xiong, Yuhao Zhou, Xiao Wang, Zhiheng Xi, Xiaoran Fan, Shiliang Pu, Jiang Zhu, Rui Zheng, Tao Gui, Qi Zhang, and Xuanjing Huang. LoRAMoE: Alleviating world knowledge forgetting in large language models via MoE-style...

    1932

  55. [63]

    Adapterfusion: Non-destructive task composition for transfer learning

    Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Non-destructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020

    2005 arXiv

  56. [64]

    Adapterswap: Continuous training of llms with data removal and access-control guarantees

    William Fleshman, Aleem Khan, Marc Marone, and Benjamin Van Durme. Adapterswap: Continuous training of llms with data removal and access-control guarantees. CoRR, 2024

    2024

  57. [65]

    Exact unlearning of finetuning data via model merging at scale

    Kevin Kuo, Amrith Setlur, Kartik Srinivas, Aditi Raghunathan, and Virginia Smith. Exact unlearning of finetuning data via model merging at scale. arXiv preprint arXiv:2504.04626, 2025

    2025 arXiv

  58. [66]

    Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models

    Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. Deepseekmoe: Towards ultimate expert specializa- tion in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024

    2024 arXiv

  59. [67]

    Generative representational instruction tuning

    Niklas Muennighoff, Hongjin SU, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. Generative representational instruction tuning. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  60. [68]

    Nexus: Specialization meets adaptability for efficiently training mixture of experts

    Nikolas Gritsch, Qizhen Zhang, Acyr Locatelli, Sara Hooker, and Ahmet Üstün. Nexus: Specialization meets adaptability for efficiently training mixture of experts. arXiv preprint arXiv:2408.15901, 2024

    2024 arXiv

  61. [69]

    Loraretriever: Input-aware lora retrieval and composition for mixed tasks in the wild

    Ziyu Zhao, Leilei Gan, Guoyin Wang, Wangchunshu Zhou, Hongxia Yang, Kun Kuang, and Fei Wu. Loraretriever: Input-aware lora retrieval and composition for mixed tasks in the wild. arXiv preprint arXiv:2402.09997, 2024

    2024 arXiv

  62. [70]

    Exploring the benefits of training expert language models over instruction tuning

    Joel Jang, Seungone Kim, Seonghyeon Ye, Doyoung Kim, Lajanugen Logeswaran, Moontae Lee, Kyungjae Lee, and Minjoon Seo. Exploring the benefits of training expert language models over instruction tuning. In International Conference on Machine Learning , pages 14702–14729. PMLR, 2023

    2023

  63. [71]

    Dam: Dynamic adapter merging for continual video qa learning

    Feng Cheng, Ziyang Wang, Yi-Lin Sung, Yan-Bo Lin, Mohit Bansal, and Gedas Bertasius. Dam: Dynamic adapter merging for continual video qa learning. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6805–6817. IEEE, 2025

    2025

  64. [72]

    Routerretriever: Routing over a mixture of expert embedding models

    Hyunji Lee, Luca Soldaini, Arman Cohan, Minjoon Seo, and Kyle Lo. Routerretriever: Routing over a mixture of expert embedding models. arXiv preprint arXiv:2409.02685, 2024

    2024 arXiv

  65. [73]

    Token-level adaptation of lora adapters for downstream task generalization

    Joshua Belofsky. Token-level adaptation of lora adapters for downstream task generalization. In Proceedings of the 2023 6th Artificial Intelligence and Cloud Computing Conference, pages 168–172, 2023

    2023

  66. [74]

    Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar

    Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan ...

    2025

  67. [75]

    Organize the web: Constructing domains enhances pre-training data curation, 2025

    Alexander Wettig, Kyle Lo, Sewon Min, Hannaneh Hajishirzi, Danqi Chen, and Luca Soldaini. Organize the web: Constructing domains enhances pre-training data curation, 2025

    2025

  68. [76]

    Starcoder: may the source be with you!, 2023

    Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, Qian Liu, Evgenii Zheltonozhskii, Terry Yue Zhuo, Thomas Wang, Olivier Dehaene, Mishig Davaadorj, Joel Lamy-Poirier, João Monteiro, ...

    2023

  69. [77]

    The stack: 3 tb of permissively licensed source code, 2022

    Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Muñoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The stack: 3 tb of permissively licensed source code, 2022

    2022

  70. [78]

    Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling Gu, Shengyi Huang, Matt Jordan, Nathan Lambert, Dustin Schwenk, Oyvind Tafjord, Taira Anderson, David Atkinson, Faeze Brahman, Christopher Clark, Pradeep Dasigi, Nouha Dziri, Mi...

    2025

  71. [79]

    olmocr: Unlocking trillions of tokens in pdfs with vision language models, 2025

    Jake Poznanski, Jon Borchardt, Jason Dunkelberger, Regan Huff, Daniel Lin, Aman Rangapur, Christopher Wilhelm, Kyle Lo, and Luca Soldaini. olmocr: Unlocking trillions of tokens in pdfs with vision language models, 2025

    2025

  72. [80]

    peS2o (Pretraining Efficiently on S2ORC) Dataset

    Luca Soldaini and Kyle Lo. peS2o (Pretraining Efficiently on S2ORC) Dataset. Technical re- port, Allen Institute for AI, 2023. ODC-By, https://github.com/allenai/pes2o

    2023

  73. [81]

    S2ORC: The semantic scholar open research corpus

    Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors,Proceedings of the 58th Annual Meeting of the Association for Computational Linguis...

    2020

  74. [82]

    Smollm2: When smol goes big – data-centric training of a small language model, 2025

    Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martín Blázquez, Guilherme Penedo, Lewis Tunstall, Andrés Marafioti, Hynek Kydlíˇcek, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo ...

    2025

  75. [83]

    Peters, Abhilasha Ravichander, Kyle Richardson, Zejiang Shen, Emma Strubell, Nishant Subramani, Oyvind Tafjord, Pete Walsh, Luke Zettlemoyer, Noah A

    Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, Valentin Hofmann, Ananya Harsh Jha, Sachin Kumar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian Magnusson, Jacob 17 Morrison, Niklas...

    2024

  76. [84]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1, 2018

    2018 arXiv

  77. [85]

    Boolq: Exploring the surprising difficulty of natural yes/no questions, 2019

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions, 2019

    2019

  78. [86]

    Khapra, Karthik Sankaranarayanan, and Sarath Chandar

    Amrita Saha, Vardaan Pahuja, Mitesh M. Khapra, Karthik Sankaranarayanan, and Sarath Chandar. Complex sequential question answering: Towards learning to converse over linked question answer pairs with a knowledge graph, 2018

    2018

  79. [87]

    Hellaswag: Can a machine really finish your sentence?, 2019

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence?, 2019

    2019

  80. [88]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In EMNLP, 2018

    2018

  81. [89]

    Piqa: Reasoning about physical commonsense in natural language, 2019

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language, 2019

    2019

  82. [90]

    Socialiqa: Commonsense reasoning about social interactions, 2019

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiqa: Commonsense reasoning about social interactions, 2019

    2019

  83. [91]

    Winogrande: An adversarial winograd schema challenge at scale

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. arXiv preprint arXiv:1907.10641, 2019

    1907 arXiv

  84. [92]

    Siva Reddy, Danqi Chen, and Christopher D. Manning. Coqa: A conversational question answering challenge, 2019

    2019

  85. [93]

    Squad: 100,000+ questions for machine comprehension of text, 2016

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text, 2016

    2016

  86. [94]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019

  87. [95]

    Weld, and Luke Zettlemoyer

    Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension, 2017

    2017

  88. [96]

    Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs, 2019

    Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs, 2019

    2019

  89. [97]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  90. [98]

    Mmlu-pro: A more robust and challenging multi-task language understanding benchmark, 2024

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. Mmlu-pro: A more robust and challenging multi-task language unders...

    2024

  91. [99]

    Agieval: A human-centric benchmark for evaluating foundation models, 2023

    Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models, 2023. 18

    2023

  92. [100]

    Le, Ed H

    Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V . Le, Ed H. Chi, Denny Zhou, and Jason Wei. Chal- lenging big-bench tasks and whether chain-of-thought can solve them, 2022

    2022

  93. [101]

    Training verifiers to solve math word problems, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021

    2021

  94. [102]

    Measuring mathematical problem solving with the math dataset, 2021

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset, 2021

    2021

  95. [103]

    Program synthesis with large language models, 2021

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. Program synthesis with large language models, 2021

    2021

  96. [104]

    Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation

    Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and LINGMING ZHANG. Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  97. [105]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott G...

    2021

  98. [106]

    Sciriff: A resource to enhance language model instruction- following over scientific literature, 2024

    David Wadden, Kejian Shi, Jacob Morrison, Aakanksha Naik, Shruti Singh, Nitzan Barzi- lay, Kyle Lo, Tom Hope, Luca Soldaini, Shannon Zejiang Shen, Doug Downey, Hannaneh Hajishirzi, and Arman Cohan. Sciriff: A resource to enhance language model instruction- following over scien...

    2024

  99. [107]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony S. Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston ...

    2024 arXiv

  100. [108]

    Extracting training data from large language models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Kather- ine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21), pa...

    2021

  101. [109]

    Extracting training data from diffusion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253–5270, 2023

    2023

  102. [110]

    Extracting memorized pieces of (copyrighted) books from open-weight language models

    A Feder Cooper, Aaron Gokaslan, Amy B Cyphert, Christopher De Sa, Mark A Lemley, Daniel E Ho, and Percy Liang. Extracting memorized pieces of (copyrighted) books from open-weight language models. arXiv preprint arXiv:2505.12546, 2025

    2025 arXiv

  103. [111]

    Choquette-Choo, Matthew Jagielski, Peter Kairouz, Sanmi Koyejo, Percy Liang, and Nicolas Papernot

    Ken Liu, Christopher A. Choquette-Choo, Matthew Jagielski, Peter Kairouz, Sanmi Koyejo, Percy Liang, and Nicolas Papernot. Language models may verbatim complete text they were not explicitly trained on. In Forty-second International Conference on Machine Learning, 2025

    2025

  104. [112]

    Arctic-embed: Scalable, efficient, and accurate text embedding models

    Luke Merrick, Danmei Xu, Gaurav Nuti, and Daniel Campos. Arctic-embed: Scalable, efficient, and accurate text embedding models. arXiv preprint arXiv:2405.05374, 2024

    2024 arXiv

  105. [113]

    The pushshift reddit dataset

    Jason Baumgartner, Savvas Zannettou, Brian Keegan, Megan Squire, and Jeremy Blackburn. The pushshift reddit dataset. In Proceedings of the international AAAI conference on web and social media, volume 14, pages 830–839, 2020

    2020

  106. [114]

    OLMo: Accelerating the science of language models

    Dirk Groeneveld, Iz Beltagy, Evan Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, Shane Arora, David Atkin- son, Russell Authur, Khyathi Chandu, Arman Cohan, Jennifer Dumas, Yanai Elazar, Yul- ing Gu, Jack Hessel, T...

    2024

  107. [115]

    Larsen, Sean Owen, and Jonathan Frankle

    Cody Blakeney, Mansheej Paul, Brett W. Larsen, Sean Owen, and Jonathan Frankle. Does your data spark joy? performance gains from domain upsampling at the end of training, 2024

    2024

  108. [116]

    The narrativeqa reading comprehension challenge, 2017

    Tomáš Koˇciský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. The narrativeqa reading comprehension challenge, 2017

    2017

  109. [117]

    Olmes: A standard for language model evaluations, 2025

    Yuling Gu, Oyvind Tafjord, Bailey Kuehl, Dany Haddad, Jesse Dodge, and Hannaneh Ha- jishirzi. Olmes: A standard for language model evaluations, 2025

    2025

  110. [118]

    Solving quantitative reasoning problems with language models, 2022

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. Solving quantitative reasoning problems with language models, 2022

    2022

  111. [119]

    "" For ablations with four datasets instead of eight, we used the following prompting

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A. Smith, and Chiyuan Zhang. Muse: Machine unlearning six-way evaluation for language models, 2024. 20 A Model Details A.1 Baseline Details Prompt-based routi...

    2024

  112. [120]

    To derive our initial dataset, we extracted each submission and concatenated it with its top-scoring, top-level comment

    Reddit data filtering A dataset of submission/comment pairs was derived from the PushShift Reddit dataset [113] (bulk dump as of March 2023) – the same dump used for Dolma Red- dit (https://huggingface.co/datasets/allenai/ dolma). To derive our initial dataset, we extracted ea...

    2023

  113. [121]

    Retrieval-based subreddit selection Dense retrieval was then used to identify academically- relevant subreddits for further filtering. We adapted search queries from MMLU test questions, and performed dense retrieval with these queries on the filtered Reddit data from Step #2,...

  114. [122]

    We defined 7 categories of question format inspired by variation observed in MMLU, and used these to construct prompts for QA text generation

    Format rewriting Finally, the data from Step #2 was input to a synthetic rewriting pipeline to generate academic QA items with coverage of diverse question formats. We defined 7 categories of question format inspired by variation observed in MMLU, and used these to construct p...

  115. [123]

    statement completion

  116. [124]

    statement truth verification

  117. [125]

    which-of-following-has-property-X

  118. [126]

    which-of-following-is-true

  119. [127]

    in-question-options

    in-question options For each format category we constructed a prompt for generating questions of that category given an input text. Below is an example prompt, for the “in-question-options” category. Prompts for other categories differ in 1) the content of the “For format ...”...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.