Pith. sign in

REVIEW 3 major objections 5 minor 93 references

Fairshare Data Pricing via Data Valuation for Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper argues that underpaying LLM data annotators is self-defeating, because in a transparent data market the price that maximizes seller profit—the buyer's maximum willingness to pay—also maximizes the buyer's long-run utility.

desk verdict The framework is worth a look, but the headline win-win result is built into the participation assumption, not demonstrated. read the letter →

arxiv 2502.00198 v4 pith:PWZ2TBK3 submitted 2025-01-31 cs.GT cs.CL

Luyang Zhang , Cathy Jiao , Beibei Li , Chenyan Xiong This is my paper · ORCID
classification cs.GTcs.CL MSC 91A6591B2691A80
keywords datapricingvaluationLLMmarketsStackelberggamesellerparticipationmaximumwillingnesstopayannotatorcompensationdynamicprogramming
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 tries to show that paying human annotators as little as possible is self-defeating for LLM builders: low prices push high-quality sellers out of the market, so the supply of training data and downstream model performance erode over time. It then proposes fairshare pricing, in which each seller charges the buyer's maximum willingness to pay, a price derived from data-valuation scores that quantify a dataset's contribution to model performance. In the paper's dynamic buyer–seller model, fairshare is the optimal long-run price for the buyer as well as the profit-maximizing price for the seller. Simulations on math, medical, and physical-reasoning tasks support the claimed win-win outcome, with better performance per dollar for buyers and higher sustained earnings for sellers.

What carries the argument

The load-bearing object is the fairshare price $p_t^* = \min\{u_t,b_t\}$, the buyer's maximum willingness to pay, computed from data-valuation scores that convert a training sample's marginal contribution into economic utility. Around that price the paper builds a Stackelberg game in which sellers move first and buyers respond, an infinite-horizon Bellman equation for the buyer's cumulative utility, and a participation function $\pi(p_t,p_t^*)$ giving the probability the seller stays active. The participation function is what turns cheap pricing into future scarcity: underpayment today lowers the chance of supply tomorrow, and the Bellman recursion makes that trade-off explicit. The empirical machinery is a family of data-valuation methods (including a lexical-similarity scorer and two influence-based estimators) that score each seller's data and set prices at the resulting willingness-to-pay values.

What would settle it

A real-market measurement of annotator retention as a function of wage would settle it: if the participation function $\pi(p,p^*)$ is flat or weakly increasing when wages fall, then exploitative pricing would not drive out supply and the model's central prediction fails. Concretely, estimate $\pi$ from a longitudinal gig-labeling platform by comparing retention and output quality across randomly assigned wage levels below and at the estimated ideal price.

Watch

Extended reading notes

Core claim

The central claim is that in a transparent LLM data market, the seller's profit-maximizing price (fairshare, $p_t^* = \min\{u_t,b_t\}$) is also the buyer's infinite-horizon optimum: with a participation response that declines when sellers are underpaid and a buyer who is sufficiently patient, any exploitative price $p_t < p_t^*$ is optimal only for finitely many periods. Lemma 2 pins the seller's optimal price to one of the buyers' maximum-willingness-to-pay breakpoints, and Lemma 3 closes the loop by showing the fairshare price maximizes the buyer's expected cumulative discounted utility in the Bellman equation. The paper further shows that the time threshold at which fairshare overtakes exploitative pricing increases as the buyer's discount factor decreases, meaning patient buyers should adopt fair pricing sooner. Empirically, the paper demonstrates these dynamics on several open LLMs and tasks, finding that fairshare sustains seller participation, yields the highest cumulative seller profit, and improves buyer utility and performance per dollar compared with reduced, random, and exploitative pricing.

Load-bearing premise

The result depends on sellers actually leaving the market, with probability that rises the further pay falls below their ideal price; if underpaid annotators stay because they have no better option, the predicted market collapse and the win-win conclusion no longer follow.

Editorial extensions

If this is right

  • Exploitative pricing is at best a short-run move: once seller participation responds to pay, it stops maximizing even the buyer's utility after finitely many periods.
  • Paying a seller the buyer's maximum willingness to pay is not charity; in the model it is the policy that maximizes the buyer's expected discounted cumulative utility.
  • Sellers earn the highest cumulative profit under fairshare, and the number of active sellers does not collapse over time.
  • Buyers that adopt fairshare get higher performance per dollar, and low-budget buyers trade short-term utility for long-term data availability.
  • The advantage of fairshare appears across different models and tasks and across multiple data-valuation methods, suggesting the mechanism is not tied to one scoring scheme.

Reading between the lines

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

  • If real annotator retention responds to pay the way the model assumes, fairshare gives platforms a simple operational rule: compute a data-valuation score, convert it to marginal utility, and post the price $\min\{u_t,b_t\}$ instead of negotiating a flat low wage.
  • The model predicts a testable asymmetry: cutting wages by a given fraction should cause measurable annotator dropout and quality decline; if retention turns out to be inelastic because workers lack alternatives, the lose-lose conclusion would fail and fairness would need regulatory support rather than market incentives.
  • The royalty variant in the paper hints that fair pricing could be implemented as a share of future model value, which would matter most for startups that cannot pay full data prices upfront, although the paper does not experimentally test that variant.
  • A natural next test is a real platform experiment with randomized wages, measuring both seller retention and downstream model quality, to estimate the participation function directly.
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

3 major / 5 minor

Summary. The paper proposes "fairshare," a data-pricing mechanism for LLM training data markets, in which prices are set according to data-valuation scores and, in a dynamic single-buyer/single-seller model, equal the buyer's maximum willingness to pay p*_t = min{u_t, b_t}. The theoretical component (Section 3) asserts that exploitative pricing (p_t < p*_t) leads to long-run market collapse (Lemma 1), while fairshare pricing maximizes both buyer utility and seller profit (Lemmas 2-3), with the crossover time depending on the discount factor (Lemma 4). The empirical component (Section 4) simulates buyer-seller dynamics for three LLMs on math, medical, and physical-reasoning tasks, and compares data-valuation methods for purchase selection and downstream performance-per-dollar. The paper concludes that fairshare yields a win-win outcome and that data-valuation-based pricing outperforms flat-rate baselines.

Significance. The paper addresses an important and timely problem—fair compensation in LLM data markets—and combines data valuation with a formal pricing model. Its strengths include a clear formal setup, explicit lemmas with proofs in the appendix, reproducible experiments on open-source models and datasets, and an informative comparison of data-valuation methods (Constant, Random, BM25, InflIP, DataInf) showing that influence-based valuation improves performance-per-dollar. These contributions are real and useful. However, the central 'win-win' claim is not supported beyond the paper's own behavioral assumption: the participation function in Section 4.2 is exactly the functional form assumed in Assumption 1, so the simulations restate rather than test the theory. In addition, Assumption 2 is internally inconsistent (it can force δ ≥ 1) in the paper's own experimental setting. If the assumptions are repaired and the claims are appropriately scoped, the framework could be a valuable step toward equitable data pricing; as it stands, the significance is conditional on assumptions that are neither validated nor always feasible.

major comments (3)
  1. [Section 3.2, Assumption 2 and Lemma 1] Assumption 2 (Eq. 7) requires δ ≥ 1/(1 + L min_t E[u_t - p*_t]). Since Lemma 3 defines p*_t = min{u_t, b_t}, we have u_t - p*_t = max{0, u_t - b_t}. In the simulation setup of Section 4.2, buyer budgets are drawn between 90% and 100% of total dataset utilities, so for many individual datasets the budget exceeds the dataset's utility and u_t - p*_t = 0; hence the minimum over t is zero and Assumption 2 forces δ ≥ 1. This contradicts the δ = 0.98 used in the experiments and renders the proof of Lemma 1 (Eqs. 29-31) inapplicable in the paper's own setting. The assumption needs to be reformulated—for example by requiring min_t E[u_t - p*_t] > 0 and δ < 1, or by replacing the lower-bound argument with a tighter bound on G that does not degenerate when the fairshare net surplus is zero.
  2. [Section 4.2, participation function and empirical circularity] The market simulation instantiates the participation rule as π(p, p*) = p/p* (Section 4.2), which is exactly the behavioral assumption (Assumption 1) that the theory requires. Under this rule, the 'exploitative' arm (10% of average utility) has π = 0.1 and must exhibit rapid seller exit, while the fairshare arm has π = 1 by construction. Thus Figures 4 and 10-12 do not provide independent evidence for the paper's 'lose-lose' and 'win-win' conclusions; they encode the assumed functional form. The real-data experiments in Section 4.3 evaluate data-valuation methods for purchase decisions and downstream model performance, but they do not measure how sellers respond to prices, so they cannot validate Assumption 1. The paper's own Limitations (Section G.1) lists running simulations with human buyers/sellers as future work, acknowledging this gap. The empirical claims in the abstract and conclusion should be scoped accordingly, or the participation function should be tested against real annotator or platform data.
  3. [Sections 3.1 and 3.2, model consistency] The dynamic model in Section 3.2 has the buyer choose the price p_t in the Bellman equation (8), whereas the market model in Section 3.1 specifies a Stackelberg game in which sellers set prices via Eq. (4). The paper never reconciles these two price-setting regimes. Lemma 3 calls p*_t the 'optimal price for the buyer,' but p*_t is also the seller's profit-maximizing price, making it unclear whether fairshare is an equilibrium outcome of the originally specified Stackelberg game or an optimal policy in a different game where the buyer makes a take-it-or-leave-it offer. This ambiguity affects the interpretation of the central 'mutually optimal' claim and should be clarified, preferably by specifying one game form for the theoretical results and the simulations.
minor comments (5)
  1. [Eq. (6), Assumption 1.1] The condition |π(p1) - π(p2)| ≥ L|p1 - p2| is a lower Lipschitz bound (an anti-Lipschitz or coercivity condition), not the standard upper Lipschitz continuity. The terminology should be adjusted to avoid confusion.
  2. [Throughout] There are numerous typographical errors: 'cummulative' for 'cumulative' in Figures 4 and 10-12; 'thorugh' in Section 2; 'and and active sampling' in Section 2; 'avaliable' near Eq. (1); 'MedAQ', 'PiQA', and 'medqaQA' in figure captions; and 'Demostrations' in Table 2's caption. A careful proofread is needed.
  3. [Section 3.1, Eq. (9)] The definition of MWP in Eq. (9) uses the budget surplus Δb_k(˜x_{k,N-1}) but the relationship between this surplus and the original budget b_k is not made explicit for buyers who already hold a baseline dataset. Clarifying this would help readers understand the pricing algorithm.
  4. [Lemma 4 and proof] The lemma statement says 't* is increasing as δ decreases,' and the proof concludes 't* also decreases' as δ increases. These statements are consistent, but the proof's final sentence is tersely worded and could be misread; it should explicitly state that t* increases when δ decreases.
  5. [Section 4.3 and Appendix F.2] The paper reports Spearman correlations between Oracle and InflIP (0.54 and 0.42 for MathQA and PIQA) but does not report confidence intervals or significance levels. Since the checklist states that the paper makes no formal statistical claims, the authors should either add such intervals or explicitly frame these numbers as descriptive.

Circularity Check

2 steps flagged · score 6.0 of 10

The market-simulation 'evidence' for lose-lose/win-win is generated by the assumed participation function π = p/p*, so the empirical demonstrations reduce by construction; the conditional theory itself is not circular.

  1. self definitional [Section 3.2, 'Exploitative Pricing: A Lose-Lose Outcome' (simulation paragraph, Figures 2a/2b)]
    "In addition to our theoretical findings, we run a simplified dynamic market simulation with one buyer and one seller, making sequential decisions. The utility of the dataset and buyer's budget varies randomly over time, and seller participation follows π(pt, p∗t) = pt/p∗t. Figures 2a and 2b provides consistent evidence that exploitative pricing causes rapid seller exit and an immediate shortage of training data."

    The simulation instantiates Assumption 1 as π(pt,p∗t)=pt/p∗t, so a seller's continuation probability is defined to be exactly the offered price divided by the ideal price. Under 'exploitative' pricing pt is a small fixed value, making π small by construction; the cumulative product over time then forces the participation curve in Figure 2b to decay to zero. The 'consistent evidence' of seller exit is therefore a direct computation of the assumed definition, not an independent empirical observation. The simulation would produce this outcome even if no real annotator ever behaved this way, so it adds no information beyond the input assumption.

  2. self definitional [Section 4.2, 'Participation Function' and 'Exploitative Pricing Leads to Lose-Lose Outcomes' (Figure 4)]
    "Seller participation probability is defined as π(pj,t, p∗j,t) = pj,t/p∗j,t for its simplicity and compliance with Assumptions 1 and 1.1. ... As a result, sellers exit the market over time, triggering a collapse in data supply."

    The market experiment uses the same assumed functional form as the theory, then chooses pricing arms so that fairshare pays p∗ (π=1), reduced pays 0.5p∗ (π=0.5), and exploitative pays a low fixed price (π near 0). The observed 'sellers exit the market' is just the cumulative product of these chosen π values over 100 time steps. Thus Figure 4's central lose-lose conclusion is encoded in the definition of the participation function rather than discovered from data. The Section 4.3 real-data experiments measure data valuation for purchase selection, not how sellers respond to prices, so they do not corroborate the participation model; the limitations section likewise lists 'running the simulation with human buyers/sellers' as future work.

full rationale

The formal theoretical results (Lemma 1 and Lemma 3) are conditional theorems: they assume Assumptions 1, 1.1, and 2, and the proofs are explicit. A conditional theorem is not circular merely because its conclusion follows from its assumptions. There is no load-bearing self-citation: the cited work by the authors (e.g., Jiao et al. 2024 for influence-function correlation) supports auxiliary valuation comparisons, not the fairshare optimality claim. The circularity lies in the empirical presentation. Both the simplified simulation in Section 3.2 and the market experiment in Section 4.2 install the participation response π = p/p* as a definition and then present the resulting seller exit as 'consistent evidence' or a demonstrated 'lose-lose outcome.' Because the seller participation probability is literally the price ratio, the fairshare arm has π=1 and the exploitative arm has π near zero by construction; Figures 2 and 4 are plots of the assumption, not validations of it. The theory is internally consistent, but the paper's empirical claim that exploitative pricing 'drives high-quality sellers out of the market' in real LLM data markets is not tested independently of the assumed behavioral rule. This is a partial circularity affecting the central empirical demonstration, so the score is 6 rather than higher; the conditional mathematics itself stands on its stated assumptions.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central claims rest mainly on modeling choices rather than on external empirical evidence: seller participation is assumed (and in simulations defined) to respond to price exactly as the theory needs; data valuation scores are assumed transparent to all; the dynamic game switches to buyer-set prices; and several simulation distributions are unspecified. The honest empirical contribution is the small-scale LLM fine-tuning comparison of valuation methods, which supports the data-selection part of the framework but not the pricing mechanism.

free parameters (3)
  • Participation function pi(p,p*) = pi(p,p*)=p/p* in Section 4.2 simulations
    Chosen for simplicity and to satisfy Assumptions 1 and 1.1; it directly encodes the result that fairshare sustains full participation while lower prices cause proportional attrition. No real seller data.
  • Discount factor delta = 0.98; robustness 0.999, 0.99, 0.98
    Chosen for the 100-step simulations in Section 4.2; Lemma 4 predicts the crossover threshold depends on delta. Not estimated.
  • Utility mapping coefficients gamma, beta = Not reported
    Section C.1 assumes u_k = gamma v_k + beta for the experiments, but gamma and beta are not given, so Equation (2) purchasing decisions cannot be exactly reconstructed.
assumptions (7)
  • domain assumption Information transparency: all participants observe each dataset's data valuation score
    Section 3.1 adopts this following [73]; in practice sellers (human annotators) would need platform-provided valuation scores, which the paper does not address.
  • domain assumption Stackelberg timing: sellers set prices first, then buyers choose datasets
    Section 3.1 defines the market as a non-cooperative Stackelberg game; the leader-follower structure is a modeling choice.
  • ad hoc to paper Assumption 1: participation probability pi is strictly increasing in price, with pi(0,p*)=0 and pi(p*,p*)=1
    Section 3.2; this is the behavioral premise driving Lemma 1 and the market-collapse conclusion. It is plausible and cited to [74,75] but not empirically validated for data annotators.
  • ad hoc to paper Assumption 1.1: participation sensitivity has a lower Lipschitz bound L
    Section 3.2; needed in the proof of Lemma 1 to obtain delta*G*L >= 1, ensuring the fair price dominates any exploitative price.
  • ad hoc to paper Assumption 2: discount factor delta >= 1/(1 + L*min_t E[u_t-p*_t])
    Section 3.2; the condition is tailored to make the proof of Lemma 1 go through. If min(u_t-p*_t)=0, the bound requires delta>=1, impossible for delta<1.
  • ad hoc to paper Buyers set prices in the dynamic model (Bellman equation (8))
    Section 3.2 switches from seller-set Stackelberg prices to a buyer-offer repeated game without explaining why the buyer controls the price; Lemma 3's buyer-optimality relies on this.
  • domain assumption Dataset utilities in simulations are drawn from an unstated distribution
    Section E.2 specifies budgets as percentages of total utilities but not how individual dataset utilities are generated; this affects all cumulative utility and profit curves.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fairshare Data Pricing via Data Valuation for Large Language Models." pith.science (2026). https://pith.science/paper/PWZ2TBK3

@misc{pith2026250200198,
  author       = {Pith},
  title        = {Pith review of: Fairshare Data Pricing via Data Valuation for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PWZ2TBK3}},
  note         = {Machine review of arXiv:2502.00198}
}
read the original abstract

Training data is the backbone of large language models (LLMs), yet today's data markets often operate under exploitative pricing -- sourcing data from marginalized groups with little pay or recognition. This paper introduces a theoretical framework for LLM data markets, modeling the strategic interactions between buyers (LLM builders) and sellers (human annotators). We begin with theoretical and empirical analysis showing how exploitative pricing drives high-quality sellers out of the market, degrading data quality and long-term model performance. Then we introduce fairshare, a pricing mechanism grounded in data valuation that quantifies each data's contribution. It aligns incentives by sustaining seller participation and optimizing utility for both buyers and sellers. Theoretically, we show that fairshare yields mutually optimal outcomes: maximizing long-term buyer utility and seller profit while sustaining market participation. Empirically when training open-source LLMs on complex NLP tasks, including math problems, medical diagnosis, and physical reasoning, fairshare boosts seller earnings and ensures a stable supply of high-quality data, while improving buyers' performance-per-dollar and long-term welfare. Our findings offer a concrete path toward fair, transparent, and economically sustainable data markets for LLM.

Figures

Figures reproduced from arXiv: 2502.00198 by the authors.

Figure 1
Figure 1. Illustration of the LLM data market, showing buyers (LLM builders) and sellers (annotators) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Buyer’s cumulative utility and seller participation under ideal and exploitative pricing [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Analysis of the buyer’s cumulative net utility as a function of the acquisition prices over [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: (1) buyer’s cumulative utilities with high- (Figure 4a) and low-budget buyer (Figure 4b), [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Purchased datasets for the buyer with high budget (Figure 5a) and low budget (Figure 5b) over 100 time periods. Pythia-1b; MedQA. Exploitative Pricing Leads to Lose-Lose Out￾comes: The exploitative pricing method, which sets uniformly low, fixed prices to reflect real￾…
Figure 6
Figure 6. Figure 6: Left column: Average price-per-sample cost of purchased data across math, medical, and physical reasoning data markets using different data valuation methods. Right, middle-columns: Buyers’ model performance versus cost, before purchasing (BP) data, and after purchasin…
Figure 7
Figure 7. Figure 7: Buyer’s accumulative utilities over time periods for [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: Left, middle-left columns: Correlation analysis between oracle and InflIP valuation. Right, middle-right columns: Performance analysis between between oracle and InflIP valuation. Influence-based methods, such as InflIP approximate the influence of a sample d on a d ′ …
Figure 9
Figure 9. Figure 9: Buyers’ model (top row: Pythia-1b, bottom row: Pythia-410m) performance and costs from [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: Analysis of (1) buyer’s cumulative utilities with high-budget buyer (Figures 10a, 10e [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Analysis of (1) buyer’s cumulative utilities with high-budget buyer (Figures 11a, 11e [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 12
Figure 12. Figure 12: Analysis of (1) buyer’s cumulative utilities with high-budget buyer (Figures 12a, 12e [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Number of purchased datasets for the buyer with high budget over time periods ( [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: Number of purchased datasets for the buyer with low budget over time periods ( [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

93 extracted references · 54 canonical work pages

  1. [1]

    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 Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    2021

  2. [2]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  3. [3]

    Galactica: A large language model for science.arXiv preprint arXiv:2211.09085, 2022

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. Galactica: A large language model for science.arXiv preprint arXiv:2211.09085, 2022

    arXiv 2022

  4. [4]

    Big tech companies in race to buy training data for artificial intelligence.Business Standard, April 2024

    Reuters. Big tech companies in race to buy training data for artificial intelligence.Business Standard, April 2024

    2024

  5. [5]

    Inside big tech’s underground race to buy ai training data.Reuters, April 5 2024

    Katie Paul and Anna Tong. Inside big tech’s underground race to buy ai training data.Reuters, April 5 2024

    2024

  6. [6]

    Wang, Haocheng Xia, Li Xiong, Xiaohui Yu, and James Zou

    Jiayao Zhang, Yuran Bi, Mengye Cheng, Jinfei Liu, Kui Ren, Qiheng Sun, Yihang Wu, Yang Cao, Raul Castro Fernandez, Haifeng Xu, Ruoxi Jia, Yongchan Kwon, Jian Pei, Jiachen T. Wang, Haocheng Xia, Li Xiong, Xiaohui Yu, and James Zou. A survey on data markets, 2024

    2024

  7. [7]

    Algorithmic impact assessments and accountability: The co-construction of impacts

    Jacob Metcalf, Emanuel Moss, Elizabeth Anne Watkins, Ranjit Singh, and Madeleine Clare Elish. Algorithmic impact assessments and accountability: The co-construction of impacts. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 735–746, 2021

    2021

  8. [8]

    The market for “lemons”: Quality uncertainty and the market mechanism

    George A Akerlof. The market for “lemons”: Quality uncertainty and the market mechanism. InUncertainty in economics, pages 235–251. Elsevier, 1978

    1978

Show all 93 references
  1. [9]

    performance of crowds

    Winter Mason and Duncan J. Watts. Financial incentives and the "performance of crowds". In Association for Computing Machinery, HCOMP ’09, page 77–85, New York, NY , USA, 2009

    2009

  2. [10]

    Kotaro Hara, Abigail Adams, Kristy Milland, Saiph Savage, Chris Callison-Burch, and Jef- frey P. Bigham. A data-driven analysis of workers’ earnings on amazon mechanical turk. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, CHI ’18, page 1–14, ...

    2018

  3. [11]

    Labelers Training AI Say They’re Overworked, Underpaid, and Exploited: 60 Minutes Transcript, November 2024

    CBS News. Labelers Training AI Say They’re Overworked, Underpaid, and Exploited: 60 Minutes Transcript, November 2024. Accessed: 2024-11-28

    2024

  4. [12]

    Openai used kenyan workers on less than $2 per hour to make chatgpt less toxic

    Time Staff. Openai used kenyan workers on less than $2 per hour to make chatgpt less toxic. https://time.com/6247678/openai-chatgpt-kenya-workers/ , 2023. Accessed: 2025-04-18

    2023

  5. [13]

    Eun Seo Jo and Timnit Gebru. Lessons from archives: Strategies for collecting sociocultural data in machine learning.Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT* 2020), pages 306–316, 2020

    2020

  6. [14]

    The impossibility of automating ambiguity

    Abeba Birhane, Vinay Uday Prabhu, Emmanuel Kahembwe, Olivia Guest, and Kirstie Whitaker Jamieson. The impossibility of automating ambiguity. InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT), pages 369–381. ACM, 2021. 11

    2021

  7. [15]

    parachute

    Lynsey Chutel. African scientists are being sidelined by “parachute” research teams. Quartz Africa, 2019

    2019

  8. [16]

    Bender, Emily Denton, and Alex Hanna

    Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research.Patterns, 2(11):100336, 2021

    2021

  9. [17]

    everyone wants to do the model work, not the data work

    Nithya Sambasivan, Anirudh Kasyap, David Thomas, Mrinmaya Natarajan, Nithin Balu, and Samir Passi Ahuja. “everyone wants to do the model work, not the data work”: Data cascades in high-stakes ai. InProceedings of the 2021 CHI Conference on Human Factors in Computing Systems (C...

    2021

  10. [18]

    Harper Business, 2019

    Mary L Gray and Siddharth Suri.Ghost work: How to stop Silicon Valley from building a new global underclass. Harper Business, 2019

    2019

  11. [19]

    Job market signaling

    Michael Spence. Job market signaling. InUncertainty in economics, pages 281–306. Elsevier, 1978

    1978

  12. [20]

    Estimating training data influence by tracing gradient descent.Advances in Neural Information Processing Systems, 33:19920–19930, 2020

    Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. Estimating training data influence by tracing gradient descent.Advances in Neural Information Processing Systems, 33:19920–19930, 2020

    2020

  13. [21]

    Datainf: Efficiently estimating data influence in lora-tuned llms and diffusion models.arXiv preprint arXiv:2310.00902, 2023

    Yongchan Kwon, Eric Wu, Kevin Wu, and James Zou. Datainf: Efficiently estimating data influence in lora-tuned llms and diffusion models.arXiv preprint arXiv:2310.00902, 2023

    2023 arXiv

  14. [22]

    Less: Selecting influential data for targeted instruction tuning, 2024

    Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen. Less: Selecting influential data for targeted instruction tuning, 2024

    2024

  15. [23]

    Mates: Model-aware data selection for efficient pretraining with data influence models, 2024

    Zichun Yu, Spandan Das, and Chenyan Xiong. Mates: Model-aware data selection for efficient pretraining with data influence models, 2024

    2024

  16. [24]

    Oxford university press New York, 1995

    Andreu Mas-Colell, Michael Dennis Whinston, Jerry R Green, et al.Microeconomic theory, volume 1. Oxford university press New York, 1995

    1995

  17. [25]

    Theory of games and economic behavior: 60th anniversary commemorative edition

    John V on Neumann and Oskar Morgenstern. Theory of games and economic behavior: 60th anniversary commemorative edition. InTheory of games and economic behavior. Princeton university press, 2007

    2007

  18. [26]

    Understanding procedural justice and its impact on business organizations

    Mary A Konovsky. Understanding procedural justice and its impact on business organizations. Journal of management, 26(3):489–511, 2000

    2000

  19. [27]

    Mathqa: Towards interpretable math word problem solving with operation-based formalisms, 2019

    Aida Amini, Saadia Gabriel, Peter Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. Mathqa: Towards interpretable math word problem solving with operation-based formalisms, 2019

    2019

  20. [28]

    What disease does this patient have? a large-scale open domain question answering dataset from medical exams.Applied Sciences, 11(14):6421, 2021

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams.Applied Sciences, 11(14):6421, 2021

    2021

  21. [29]

    Piqa: Reasoning about phys- ical commonsense in natural language

    Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about phys- ical commonsense in natural language. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432–7439, 2020

    2020

  22. [30]

    Improvements to bm25 and language models examined

    Andrew Trotman, Antti Puurula, and Blake Burgess. Improvements to bm25 and language models examined. InProceedings of the 19th Australasian Document Computing Symposium, pages 58–65, 2014

    2014

  23. [31]

    A marketplace for data: An algorithmic solution

    Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. A marketplace for data: An algorithmic solution. InProceedings of the 2019 ACM Conference on Economics and Computation, pages 701–726, 2019

    2019

  24. [32]

    Towards model-based pricing for ma- chine learning in a data marketplace

    Lingjiao Chen, Paraschos Koutris, and Arun Kumar. Towards model-based pricing for ma- chine learning in a data marketplace. InProceedings of the 2019 international conference on management of data, pages 1535–1552, 2019. 12

    2019

  25. [33]

    Selling data to a machine learner: Pricing via costly signaling

    Junjie Chen, Minming Li, and Haifeng Xu. Selling data to a machine learner: Pricing via costly signaling. InInternational Conference on Machine Learning, pages 3336–3359. PMLR, 2022

    2022

  26. [34]

    Selling consumer data for profit: Optimal market-segmentation design and its consequences.American Economic Review, 112(4):1364–1393, 2022

    Kai Hao Yang. Selling consumer data for profit: Optimal market-segmentation design and its consequences.American Economic Review, 112(4):1364–1393, 2022

    2022

  27. [35]

    A survey on data markets.arXiv preprint arXiv:2411.07267, 2024

    Jiayao Zhang, Yuran Bi, Mengye Cheng, Jinfei Liu, Kui Ren, Qiheng Sun, Yihang Wu, Yang Cao, Raul Castro Fernandez, Haifeng Xu, et al. A survey on data markets.arXiv preprint arXiv:2411.07267, 2024

    2024 arXiv

  28. [36]

    A distributional framework for data valuation

    Amirata Ghorbani, Michael Kim, and James Zou. A distributional framework for data valuation. InInternational Conference on Machine Learning, pages 3535–3544. PMLR, 2020

    2020

  29. [37]

    Quantifying the invisible labor in crowd work

    Carlos Toxtli, Siddharth Suri, and Saiph Savage. Quantifying the invisible labor in crowd work. Proceedings of the ACM on human-computer interaction, 5(CSCW2):1–26, 2021

    2021

  30. [38]

    Ai boom relies on underpaid workers in the philippines.The Washington Post, 2023

    Drew Harwell and Regine Cabato. Ai boom relies on underpaid workers in the philippines.The Washington Post, 2023. Accessed: 2025-04-20

    2023

  31. [39]

    Ethical and social risks of harm from language models, 2021

    Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks...

    2021

  32. [40]

    AI Training Data Has a Price Tag That Only Big Tech Can Afford, June 2024

    TechCrunch. AI Training Data Has a Price Tag That Only Big Tech Can Afford, June 2024. Accessed: 2024-11-28

    2024

  33. [41]

    The next great leap in ai is behind schedule and crazy expensive.The Wall Street Journal, December 2024

    Deepa Seetharaman. The next great leap in ai is behind schedule and crazy expensive.The Wall Street Journal, December 2024

    2024

  34. [42]

    Why amazon web services ceo matt garman is playing the long game on ai

    Ayesha Javed. Why amazon web services ceo matt garman is playing the long game on ai. TIME, February 2025

    2025

  35. [43]

    Export subsidies and international market share rivalry

    James A Brander and Barbara J Spencer. Export subsidies and international market share rivalry. Journal of international Economics, 18(1-2):83–100, 1985

    1985

  36. [44]

    Springer Science & Business Media, 2013

    Jonathan F Bard.Practical bilevel optimization: algorithms and applications, volume 30. Springer Science & Business Media, 2013

    2013

  37. [45]

    Springer Science & Business Media, 2010

    Heinrich V on Stackelberg.Market structure and equilibrium. Springer Science & Business Media, 2010

    2010

  38. [46]

    A survey on data pricing: from economics to data science.IEEE Transactions on knowledge and Data Engineering, 34(10):4586–4608, 2020

    Jian Pei. A survey on data pricing: from economics to data science.IEEE Transactions on knowledge and Data Engineering, 34(10):4586–4608, 2020

    2020

  39. [47]

    A game theory approach in seller–buyer supply chain.European Journal of Operational Research, 195(2):442– 448, 2009

    Maryam Esmaeili, Mir-Bahador Aryanezhad, and Panlop Zeephongsekul. A game theory approach in seller–buyer supply chain.European Journal of Operational Research, 195(2):442– 448, 2009

    2009

  40. [48]

    Dependable de- mand response management in the smart grid: A stackelberg game approach.IEEE Transactions on Smart Grid, 4(1):120–132, 2013

    Sabita Maharjan, Quanyan Zhu, Yan Zhang, Stein Gjessing, and Tamer Basar. Dependable de- mand response management in the smart grid: A stackelberg game approach.IEEE Transactions on Smart Grid, 4(1):120–132, 2013

    2013

  41. [49]

    Harvard University Press, 1989

    Nancy L Stokey and Robert E Lucas Jr.Recursive methods in economic dynamics. Harvard University Press, 1989

    1989

  42. [50]

    Sage, 1998

    Robert G Folger, Robert Folger, and Russell Cropanzano.Organizational justice and human resource management, volume 7. Sage, 1998

    1998

  43. [51]

    Fairness theory: Justice as accountability.Advances in organizational justice, 1(1-55):12, 2001

    Robert Folger and Russell Cropanzano. Fairness theory: Justice as accountability.Advances in organizational justice, 1(1-55):12, 2001. 13

    2001

  44. [52]

    What is organizational justice? a historical overview

    Jason A Colquitt, Jerald Greenberg, and Cindy P Zapata-Phelan. What is organizational justice? a historical overview. InHandbook of organizational justice, pages 3–56. Psychology Press, 2013

    2013

  45. [53]

    Organizational justice research: A review, synthesis, and research agenda

    Mladen Adamovic. Organizational justice research: A review, synthesis, and research agenda. European Management Review, 20(4):762–782, 2023

    2023

  46. [54]

    Training data influence analysis and estimation: A survey

    Zayd Hammoudeh and Daniel Lowd. Training data influence analysis and estimation: A survey. Machine Learning, 113(5):2351–2403, 2024

    2024

  47. [55]

    Understanding black-box predictions via influence functions

    Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. InInternational conference on machine learning, pages 1885–1894. PMLR, 2017

    2017

  48. [56]

    Estimating training data influence by tracing gradient descent, 2020

    Garima Pruthi, Frederick Liu, Mukund Sundararajan, and Satyen Kale. Estimating training data influence by tracing gradient descent, 2020

    2020

  49. [57]

    Understanding instance-level impact of fairness constraints

    Jialu Wang, Xin Eric Wang, and Yang Liu. Understanding instance-level impact of fairness constraints. InInternational Conference on Machine Learning, pages 23114–23130. PMLR, 2022

    2022

  50. [58]

    Ma, and Chenyan Xiong

    Cathy Jiao, Yijun Pan, Emily Xiao, Daisy Sheng, Niket Jain, Hanzhang Zhao, Ishita Dasgupta, Jiaqi W. Ma, and Chenyan Xiong. Date-lm: Benchmarking data attribution evaluation for large language models, 2025

    2025

  51. [59]

    Unlearning traces the influential training data of language models.arXiv preprint arXiv:2401.15241, 2024

    Masaru Isonuma and Ivan Titov. Unlearning traces the influential training data of language models.arXiv preprint arXiv:2401.15241, 2024

    2024 arXiv

  52. [60]

    On influence functions, classification influence, relative influence, memorization and generalization.arXiv preprint arXiv:2305.16094, 2023

    Michael Kounavis, Ousmane Dia, and Ilqar Ramazanli. On influence functions, classification influence, relative influence, memorization and generalization.arXiv preprint arXiv:2305.16094, 2023

    2023 arXiv

  53. [61]

    Hint: Healthy influential-noise based training to defend against data poisoning attacks

    Minh-Hao Van, Alycia N Carey, and Xintao Wu. Hint: Healthy influential-noise based training to defend against data poisoning attacks. In2023 IEEE International Conference on Data Mining (ICDM), pages 608–617. IEEE, 2023

    2023

  54. [62]

    Fairness without harm: An influence-guided active sampling approach

    Jinlong Pang, Jialu Wang, Zhaowei Zhu, Yuanshun Yao, Chen Qian, and Yang Liu. Fairness without harm: An influence-guided active sampling approach. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  55. [63]

    A value for n-person games.Contribution to the Theory of Games, 2, 1953

    Lloyd S Shapley. A value for n-person games.Contribution to the Theory of Games, 2, 1953

    1953

  56. [64]

    Towards efficient data valuation based on the shapley value

    Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J Spanos. Towards efficient data valuation based on the shapley value. InThe 22nd International Conference on Artificial Intelligence and Statistics, pag...

    2019

  57. [65]

    Data shapley: Equitable valuation of data for machine learning

    Amirata Ghorbani and James Zou. Data shapley: Equitable valuation of data for machine learning. InInternational conference on machine learning, pages 2242–2251. PMLR, 2019

    2019

  58. [66]

    Cs-shapley: class-wise shapley values for data valuation in classification.Advances in Neural Information Processing Systems, 35:34574– 34585, 2022

    Stephanie Schoch, Haifeng Xu, and Yangfeng Ji. Cs-shapley: class-wise shapley values for data valuation in classification.Advances in Neural Information Processing Systems, 35:34574– 34585, 2022

    2022

  59. [67]

    Data shapley in one training run

    Jiachen T Wang, Prateek Mittal, Dawn Song, and Ruoxi Jia. Data shapley in one training run. arXiv preprint arXiv:2406.11011, 2024

    2024 arXiv

  60. [68]

    Du-shapley: A shapley value proxy for efficient dataset valuation.Advances in Neural Information Processing Systems, 37:1973–2000, 2024

    Felipe Garrido Lucero, Benjamin Heymann, Maxime V ono, Patrick Loiseau, and Vianney Perchet. Du-shapley: A shapley value proxy for efficient dataset valuation.Advances in Neural Information Processing Systems, 37:1973–2000, 2024

    1973

  61. [69]

    Towards tracing knowledge in language models back to the training data

    Ekin Akyurek, Tolga Bolukbasi, Frederick Liu, Binbin Xiong, Ian Tenney, Jacob Andreas, and Kelvin Guu. Towards tracing knowledge in language models back to the training data. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors,Findings of the Association for Computatio...

    2022

  62. [70]

    Enhancing training data attribution for large language models with fitting error consideration

    Kangxi Wu, Liang Pang, Huawei Shen, and Xueqi Cheng. Enhancing training data attribution for large language models with fitting error consideration. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural...

    2024

  63. [71]

    In-context probing approximates influence function for data valuation.arXiv preprint arXiv:2407.12259, 2024

    Cathy Jiao, Gary Gao, and Chenyan Xiong. In-context probing approximates influence function for data valuation.arXiv preprint arXiv:2407.12259, 2024

    2024 arXiv

  64. [72]

    Greats: Online selection of high-quality data for llm training in every iteration.Advances in Neural Information Processing Systems, 37:131197–131223, 2024

    Jiachen Tianhao Wang, Tong Wu, Dawn Song, Prateek Mittal, and Ruoxi Jia. Greats: Online selection of high-quality data for llm training in every iteration.Advances in Neural Information Processing Systems, 37:131197–131223, 2024

    2024

  65. [73]

    Reputation and feedback systems in online platform markets.Annual review of economics, 8(1):321–340, 2016

    Steven Tadelis. Reputation and feedback systems in online platform markets.Annual review of economics, 8(1):321–340, 2016

    2016

  66. [74]

    Bottom-fishing and declining prices in sequential auctions

    Ian L Gale and Donald B Hausch. Bottom-fishing and declining prices in sequential auctions. Games and Economic Behavior, 7(3):318–331, 1994

    1994

  67. [75]

    Sequential bargaining in the field: Evidence from millions of online bargaining interactions.The Quarterly Journal of Economics, 135(3):1319–1361, 2020

    Matthew Backus, Thomas Blake, Brad Larsen, and Steven Tadelis. Sequential bargaining in the field: Evidence from millions of online bargaining interactions.The Quarterly Journal of Economics, 135(3):1319–1361, 2020

    2020

  68. [76]

    Andrew R. Chow. Why amazon web services ceo matt garman is playing the long game on ai. Time, 2024

    2024

  69. [77]

    Openai’s next big ai effort, gpt-5, is behind schedule and crazy expensive.The Wall Street Journal, 2024

    Tom Dotan. Openai’s next big ai effort, gpt-5, is behind schedule and crazy expensive.The Wall Street Journal, 2024

    2024

  70. [78]

    Sundar pichai: The 100 most influential people in ai 2024.Time, 2024

    Jane Li. Sundar pichai: The 100 most influential people in ai 2024.Time, 2024

    2024

  71. [79]

    Superagency in the workplace: Empowering people to unlock ai’s full potential

    Hannah Mayer, Lareina Yee, Michael Chui, and Roger Roberts. Superagency in the workplace: Empowering people to unlock ai’s full potential. https://www.mckinsey.com/capabilit ies/mckinsey-digital/our-insights/superagency-in-the-workplace-empower ing-people-to-unlock-ais-full-po...

    2025

  72. [80]

    Dynamic programming and statistical communication theory.Proceedings of the National Academy of Sciences, 43(8):749–751, 1957

    Richard Bellman and Robert Kalaba. Dynamic programming and statistical communication theory.Proceedings of the National Academy of Sciences, 43(8):749–751, 1957

    1957

  73. [81]

    The llama 3 herd of models, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and etc. The llama 3 herd of models, 2024

    2024

  74. [82]

    Pythia: A suite for analyzing large language models across training and scaling

    Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In Intern...

    2023

  75. [83]

    Fine-tuning is fine, if calibrated

    Zheda Mai, Arpita Chowdhury, Ping Zhang, Cheng-Hao Tu, Hong-You Chen, Vardaan Pahuja, Tanya Berger-Wolf, Song Gao, Charles Stewart, Yu Su, et al. Fine-tuning is fine, if calibrated. arXiv preprint arXiv:2409.16223, 2024

    2024 arXiv

  76. [84]

    Training verifiers to solve math word problems, 2021.URL https://arxiv

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems, 2021.URL https://arxiv. org/abs/2110.14168, 2021

    2021 arXiv

  77. [85]

    A survey of deep learning for mathematical reasoning

    Pan Lu, Liang Qiu, Wenhao Yu, Sean Welleck, and Kai-Wei Chang. A survey of deep learning for mathematical reasoning. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume...

    2023

  78. [86]

    Mathematical language models: A survey.arXiv preprint arXiv:2312.07622, 2023

    Wentao Liu, Hanglei Hu, Jie Zhou, Yuyang Ding, Junsong Li, Jiayi Zeng, Mengliang He, Qin Chen, Bo Jiang, Aimin Zhou, et al. Mathematical language models: A survey.arXiv preprint arXiv:2312.07622, 2023. 15

    2023 arXiv

  79. [87]

    Large language models for mathematical reasoning: Progresses and challenges

    Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. Large language models for mathematical reasoning: Progresses and challenges. In Neele Falk, Sara Papi, and Mike Zhang, editors,Proceedings of the 18th Conference of the European Chapter of the Association ...

    2024

  80. [88]

    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. InInternational Conference on Learning Representations, 2021

    2021

  81. [89]

    A survey of large language models in medicine: Progress, application, and challenge.arXiv preprint arXiv:2311.05112, 2023

    Hongjian Zhou, Fenglin Liu, Boyang Gu, Xinyu Zou, Jinfa Huang, Jinge Wu, Yiru Li, Sam S Chen, Peilin Zhou, Junling Liu, et al. A survey of large language models in medicine: Progress, application, and challenge.arXiv preprint arXiv:2311.05112, 2023

    2023 arXiv

  82. [90]

    Ed-copilot: Reduce emergency department wait time with language model diagnostic assistance, 2024

    Liwen Sun, Abhineet Agarwal, Aaron Kornblith, Bin Yu, and Chenyan Xiong. Ed-copilot: Reduce emergency department wait time with language model diagnostic assistance, 2024

    2024

  83. [91]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2021. 16 A Influence-based Data Valuation In Section 4, we introduced...

    2021

  84. [92]

    upweighting

    All models are trained on A6000 GPUs on single GPU settings and take less than 1 hour. Model Evaluation:For evaluation, we use the test splits of the previously mentioned datasets. In particular, we use 5-shot evaluation on the MathQA test set, and 4-shot evaluation in on the ...

  85. [93]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.