Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Gradients: When Markets Meet Fine-tuning -- A Distributed Approach to Model Optimisation

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Rewarding independent miners for results finds better fine-tuning configurations than centralized AutoML in controlled trials.

desk verdict Novel marketplace for fine-tuning, but the headline win rates are confounded by an unmatched 8-25x exploration budget. read the letter →

arxiv 2506.07940 v2 pith:7LCQFVAE submitted 2025-06-09 cs.AI cs.LG

Christopher Subia-Waud (Rayonlabs Team) This is my paper
classification cs.AIcs.LG
keywords fine-tuningAutoMLhyperparameteroptimizationdecentralizedmachinelearningincentivedesigncompetitivetournamenttheorydiffusionmodels
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 argues that the best way to optimize a fine-tuning run is not to run one search algorithm but to let many independent solvers compete for a reward. It reports that such a market-based system, tested on 180 fine-tuning tasks from 70M to 70B parameter models, beat every centralized AutoML platform it was compared with: a perfect record against three commercial cloud services and an 82.8% win rate against the leading no-code AutoML service. The mean loss improvements were 11.8% against the no-code service and 42.1% against the commercial clouds, with task-specific gains of 30–40% on retrieval and reasoning. If the comparison holds, the conclusion is that economic incentives can drive fine-tuning exploration more effectively than algorithmic search.

What carries the argument

The load-bearing mechanism is a two-sided tournament. Validators define tasks and evaluate submitted models in isolated environments on weighted losses; miners are selected with probability weighted by past quality, and the reward function is a steep sigmoid of the normalized score, so first place is worth far more than the rest and bottom-ranked miners are penalized. That reward asymmetry turns each task into a contest in which every miner must outdo the others, which the paper argues pushes exploration into unusual corners of hyperparameter space and keeps strategies diverse because winners do not disclose their methods.

What would settle it

Give a centralized AutoML platform the same aggregate parallel budget (for example, 8–25 full training runs per task) on the same datasets and metrics; if it matches or beats the tournament's losses, the claim that market incentives rather than raw search budget produce the gains is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that competition itself, not a better optimizer, is the mechanism that finds superior configurations. In the Gradients system, validators specify a task and 8–25 miners independently train models using whatever hyperparameters they choose; the results are evaluated on held-out and synthetic data, and the best-scoring miners receive a disproportionately large share of token rewards. Under that incentive scheme, the system reports lower losses than centralized platforms across nearly all of 180 controlled experiments, with the gap widening on harder tasks such as retrieval-augmented generation and high-denoise image reconstruction. The authors conclude that tournament-driven exploration explores regions of configuration space that centralized Bayesian or grid search never reaches.

Load-bearing premise

The central result assumes that the comparison with centralized services is fair even though the market system spends 8 to 25 times more compute per task, meaning the reported advantage is attributed to competition rather than to a larger exploration budget.

Editorial extensions

If this is right

  • If the result is correct, users of fine-tuning services can obtain the same model quality from the same data by outsourcing the search to competing independent solvers, without needing to know which hyperparameters work.
  • It would mean that centralized AutoML systems, which optimize a single search trajectory, leave a measurable performance reserve on the table, especially on tasks with complex reasoning or long-context grounding.
  • For diffusion fine-tuning, the paper's reported advantage at high noise levels implies that competitive search finds configurations that are not merely better on average but more robust under degraded conditions.
  • The reported need for 8–25 parallel runs per task means that the economic advantage comes with a computational and energy cost, so the method is most worthwhile when application performance, not efficiency, is the priority.

Reading between the lines

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

  • Editorial inference: the single most decisive confound is compute budget; if a centralized platform were allowed the same 8–25 parallel runs per task, the gap could be largely closed, which would reclassify the finding from 'market beats planning' to 'parallel search beats single search.'
  • Editorial inference: the paper's observation that miners use low-rank adapters with ranks far above theoretical recommendations suggests that competitive search is discovering useful empirical practices that theory does not yet justify; those discovered configurations could be harvested as heuristics even if the market is not retained.
  • Editorial inference: the claim that strategic opacity is essential predicts a testable consequence: a version of the system that reveals winning configurations after each round should lose its diversity advantage and show shrinking performance gains over time.
  • Editorial inference: because the reported gains are largest on RAG and reasoning tasks, the market mechanism may be especially valuable as an automated way of matching fine-tuning schedules to data distribution shifts, a connection the paper does not explore.
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 / 6 minor

Summary. This paper presents Gradients, a decentralized fine-tuning marketplace built on the Bittensor network. Validators create fine-tuning tasks; miners independently choose hyperparameters, LoRA configurations, and optimization strategies; the system evaluates submitted models on held-out test and synthetic sets and distributes token rewards based on relative performance. The authors report 180 controlled experiments comparing Gradients with HuggingFace AutoTrain, TogetherAI, Databricks, Google Cloud, and CivitAI, claiming win rates of 82.8% against AutoTrain and 100% against the commercial platforms, with mean improvements up to 42.1%. They interpret these results as evidence that economic incentives and competition systematically outperform centralized AutoML.

Significance. If the central empirical claim were supported, this would be a notable contribution: it would provide evidence that competitive, incentive-driven search can outperform centralized AutoML across models from 70M to 70B parameters and diverse task types. The paper has real strengths: the system is implemented and deployed, the code is released, the experimental breadth is substantial, and the authors explicitly acknowledge the resource trade-off in Section 6. However, the central comparison is confounded: Gradients uses 8–25 parallel training runs per task while baselines are reported as single runs or small searches, so the reported win rates may reflect best-of-N exploration rather than market incentives. Because the paper does not provide per-task data, statistical tests, or a matched-budget baseline, the strength of the conclusion currently exceeds what the evidence supports. With additional controls or reanalysis, the contribution could be valuable.

major comments (3)
  1. [§3.7, §4.1, §4.3, §6] The comparison does not hold the number of training runs or total compute constant. Section 3.7 states that 8–15 miners are assigned to text tasks and 15–25 to image tasks, and Section 4.3 repeats this protocol. Section 4.1 claims “identical conditions” across platforms, but the list of controls includes only model, data, training time, and evaluation metric, not compute or the number of trials. The Discussion (“Computational Trade-offs”) explicitly concedes that each task requires 8–25 parallel training runs versus single runs in centralized systems. Under noisy fine-tuning outcomes, taking the best of N independent runs has a purely statistical advantage over a single run, so the 82.8% and 100% win rates reported in Section 5 do not by themselves establish that competition or reward design causes the gains. The paper needs an equal-budget control (e.g., random search or multi-start with the same number of runs) or an analysis relating win rates to miner pool size.
  2. [§5, Figure 2] The headline aggregate win rates and mean improvements are presented without per-task results, standard errors, confidence intervals, or significance tests. Section 5.1 acknowledges that the number of 70B-model experiments was limited and may affect statistical significance, but the main claims aggregate over heterogeneous tasks, models, and platforms. Without raw per-task numbers or statistical analysis, the reader cannot assess whether the 11.8% and 42.1% mean improvements are consistent effects or driven by a few outliers, and the claim of “consistent superiority” is not supported as stated.
  3. [§4.5] The baseline comparisons are not sufficiently characterized. The paper lists HuggingFace AutoTrain, TogetherAI, Databricks, Google Cloud Vertex AI, and CivitAI, but it does not describe the hyperparameters used, the number of trials run by each baseline, the hardware, or the procedure for selecting the baseline result. Without evidence that each baseline represents that platform’s best achievable configuration, the 100% win rates against commercial platforms may reflect weak baseline settings rather than Gradients’ superiority. This is particularly important because the baselines are commercial services whose default or standard configurations are not necessarily their optimal ones.
minor comments (6)
  1. [§3.5, Eq. (12)] The symbol ω is reused in Eq. (12) for the linear weight in the final score, after already being used in Section 3.4 for the test-score weighting and in Eq. (11) for temporal-window weights; this makes the notation unnecessarily confusing.
  2. [§3.6] The quantity V_trust in the on-chain weight formula W_chain(m_i) = S_final(m_i) · V_trust is never defined in the paper.
  3. [§3.3, §4.3] The dataset split parameters in Section 3.3 set ρ_synth = 1.0 and κ_synth = 300, which is inconsistent with the statement in Section 4.3 that splits are “typically 80/10/synth”; the relationship between the formal parameters and the typical split should be clarified.
  4. [§4.5, §5] CivitAI is introduced as a diffusion baseline in Section 4.5, but the reported win-rate matrix in Section 5 (Figure 2) covers only HuggingFace, TogetherAI, Databricks, and Google Cloud; the diffusion improvements in the abstract and conclusion are not tied to a named comparator, so the reader cannot determine what the 23.4% figure refers to.
  5. [§4.6, §4.7, §5] The task lists in Sections 4.6 and 4.7 do not give the number of experiments per task category, even though the aggregate 180-experiment count and the categorical win rates in Figure 1 depend on this balance; providing the breakdown would help the reader evaluate whether the aggregate results are dominated by particular task types.
  6. [§5.1] The statement that 70B experiments were limited due to computational cost is important and should be reflected in the abstract and conclusion, which currently state the win rates and improvements without this caveat.

Circularity Check

1 steps flagged · score 4.0 of 10

The reported win rate is partly a best-of-N selection artifact because the loss used to rank miners is the same loss later reported as the outcome; the market-incentive claim is not independently demonstrated.

  1. fitted input called prediction [Section 3.4-3.5 (Eqs. 5-9) vs Section 4.4.1]
    "Within each task, miners are ranked by their weighted loss, with scores assigned according to: ... For text models, we computed direct loss on held-out test sets."

    Eq. 5 (L_test) and Eq. 6 (L_synth) define L_weighted, and Eq. 9 ranks miners by that L_weighted. The reported head-to-head results use the same held-out test loss (Section 4.4.1: 'computed direct loss on held-out test sets'). Thus the selected winning miner is the one that minimizes the very quantity later reported as Gradients' performance, so the win rate is partly a restatement of the selection rule. The effect is statistically amplified by selecting the best of 8-25 parallel runs versus single runs in the centralized baselines, which the Discussion explicitly concedes. This does not make the benchmark fully tautological, but it removes the independence of the evaluation from the selection objective.

full rationale

This is primarily an empirical benchmark paper rather than a derivation, so there is no equation-level cycle in the usual sense: the reward equations do not mathematically force the reported 82.8% and 100% win rates, and the comparison against external platforms involves real data collection. The main circularity-adjacent issue is that the loss used for miner ranking (L_weighted, built from L_test and L_synth) is the same loss later reported as the outcome; the winner is therefore selected on the evaluation metric, and the reported advantage is partly a best-of-N artifact from using 8-25 parallel training runs versus single runs. This weakens attribution to 'market incentives' but does not make the result entirely tautological, since the external baselines could in principle still win. The compute-budget confound is a correctness and validity concern rather than a circularity, and the Bittensor/Yuma citations are background references, not load-bearing self-citations; no uniqueness theorem or ansatz is imported from the authors' prior work. The score reflects the partial self-reference in the evaluation design, not the resource trade-off.

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

The central empirical claim rests on hand-chosen evaluation constants, including omega, delta, split ratios, and generation temperatures, and on several unverified behavioral assumptions about miners and baselines. No free parameter is fitted to data in a derivation sense, but the constants directly shape the reported numbers.

free parameters (4)
  • Text evaluation weighting omega = 0.7 (chosen, not fitted)
    Weights test loss versus synthetic loss in the miner scoring and platform comparison; changing it changes the reported win rates.
  • Image evaluation weighting delta = 0.25 (chosen, not fitted)
    Weights text-guided versus non-text-guided reconstruction loss for diffusion models; directly shapes the reported diffusion improvements.
  • Test and synthetic split ratios = rho_test=0.1, rho_synth=1.0 with caps 1000 and 300
    Determines which examples become the held-out test set and the synthetic evaluation set; affects all reported losses.
  • Synthetic generation temperatures = tau=0.6 for text, 0.4 for images
    Controls variability of generated evaluation examples; the synthetic set contributes to the weighted loss.
assumptions (5)
  • domain assumption Miners respond rationally to token rewards by searching harder and more creatively.
    This is the central mechanism of the system; no behavioral evidence is provided beyond anecdotal observations in Section 6.
  • domain assumption Held-out loss and synthetic reconstruction loss are valid proxies for real fine-tuning quality.
    Used in Section 3.4 to score miners and in Section 4.4 to compare platforms; no human evaluation or external benchmarks are reported.
  • domain assumption Commercial baseline configurations represent those platforms' best achievable performance.
    Section 4.5 describes baselines as platforms, but does not show they were tuned or given the same search budget as Gradients.
  • domain assumption The extra 8 to 25 parallel training runs used by Gradients do not explain the reported gains.
    Section 3.7 and the Discussion admit the compute asymmetry; no control experiment isolates competition from compute.
  • domain assumption L2 pixel reconstruction loss measures meaningful image-generation quality.
    Equations 14 and 15 use L2 loss only; no FID, CLIP, or human preference scores are reported for diffusion outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gradients: When Markets Meet Fine-tuning -- A Distributed Approach to Model Optimisation." pith.science (2026). https://pith.science/paper/7LCQFVAE

@misc{pith2026250607940,
  author       = {Pith},
  title        = {Pith review of: Gradients: When Markets Meet Fine-tuning -- A Distributed Approach to Model Optimisation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LCQFVAE}},
  note         = {Machine review of arXiv:2506.07940}
}
read the original abstract

Current AutoML platforms leave substantial performance untapped. Testing 180 fine-tuning tasks across models from 70M to 70B parameters, we found that HuggingFace AutoTrain, TogetherAI, Databricks, and Google Cloud consistently produce suboptimal configurations. Gradients, built on the Bittensor network, attacks this problem through competition. Independent miners race to find optimal hyperparameters, earning rewards proportional to their models' performance. This tournament drives exploration of configuration spaces that single-strategy methods never examine. In our experiments, Gradients achieved a 100\% win rate against TogetherAI, Databricks, and Google Cloud, and beat HuggingFace AutoTrain in 82.8\% of experiments. Mean improvements reached 42.1\% against commercial platforms. Retrieval-augmented generation tasks saw 30-40\% gains; diffusion models improved 23.4\% on person-specific generation. When miners compete for rewards, they develop optimization strategies that centralized approaches overlook. These findings demonstrate that decentralized systems with economic incentives can systematically outperform traditional AutoML, suggesting market dynamics may be key to achieving superior fine-tuning results. Code is available at https://github.com/rayonlabs/G.O.D.

Figures

Figures reproduced from arXiv: 2506.07940 by the authors.

Figure 1
Figure 1. Performance improvement by task type. Four-panel analysis: box plots show improvement distributions, bar charts [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Win rates against leading AutoML platforms. Matrix [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Performance magnitude analysis. (a) Improvement [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Scale and architecture effects. (a) Performance im [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 26 canonical work pages

  1. [1]

    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 . 2023. Pythia: A Suite for Ana- lyzing Large Language Models Across Training and Scaling. arXiv preprint arXiv:2304.01373 (2023)

    arXiv 2023

  2. [2]

    Yung-Chieh Chan, George Pu, Apoorv Shanker, Navdeep Jaitly, Joshua Susskind, and Mohammad Norouzi. 2024. Balancing Cost and Effectiveness of Synthetic Data Generation Strategies for LLMs. arXiv preprint arXiv:2410.23940 (2024)

    work page Pith review arXiv 2024

  3. [3]

    Sahil Chaudhary. 2023. Code Alpaca: An Instruction-following LLaMA model for code generation. https://github.com/sahil280114/codealpaca

    2023

  4. [4]

    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. 2021. Training Verifiers to Solve Math Word Problems. arXiv preprint arXiv:2110.14168 (2021)

    arXiv 2021

  5. [5]

    CodeParrot. 2023. Self-instruct-starcoder: Instruction dataset generated from StarCoder. https://huggingface.co/datasets/codeparrot/self-instruct-starcoder

    work page 2023

  6. [6]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, et al. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv preprint arXiv:2501.12948 (2025)

    arXiv 2025

  7. [7]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient Finetuning of Quantized LLMs. In Advances in Neural Informa- tion Processing Systems

    work page 2023

  8. [8]

    Ning Ding, Zhenghao Han, Sheng Zhou, and Zhisheng Niu. 2020. Incentive Mechanism Design for Distributed Coded Machine Learning. arXiv preprint arXiv:2012.08715 (2020)

    work page Pith review arXiv 2020

Show all 43 references
  1. [9]

    Zeyu Ding, Xipeng Qiu, Xuanjing Zhang, and Xuanjing Huang. 2024. Parameter- Efficient Fine-Tuning for Large Models: A Comprehensive Survey.arXiv preprint arXiv:2403.14608 (2024)

    2024 arXiv

  2. [10]

    Arkadiy Dushatskiy, Hector Mendoza, Madalina M Drugan, and Bas van Stein

  3. [11]

    Parkes, and Sai Sri- vatsa Ravindranath

    Paul Dütting, Zhe Feng, Harikrishna Narasimhan, David C. Parkes, and Sai Sri- vatsa Ravindranath. 2019. Optimal Auctions through Deep Learning. In Pro- ceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). PMLR, ...

    2019

  4. [12]

    Stefan Falkner, Aaron Klein, and Frank Hutter. 2018. BOHB: Robust and Effi- cient Hyperparameter Optimization at Scale. In Proceedings of the 35th Interna- tional Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80). PMLR, 1437–1446

    2018

  5. [13]

    Matthias Feurer, Aaron Klein, Katharina Eggensperger, Jost Springenberg, Manuel Blum, and Frank Hutter. 2015. Efficient and Robust Automated Ma- chine Learning. In Advances in Neural Information Processing Systems , Vol. 28. 2962–2970

    2015

  6. [14]

    FreedomIntelligence. 2024. RAG-Instruct: Knowledge-grounded instruction tun- ing dataset. https://huggingface.co/datasets/FreedomIntelligence/RAG-Instruct

    2024

  7. [15]

    Robert Friel, Masha Belyi, and Christophe Baik. 2024. RAGBench: Explain- able Benchmark for Retrieval-Augmented Generation Systems. arXiv preprint arXiv:2407.11005 (2024)

    2024 arXiv

  8. [16]

    Aaron Grattafiori et al . 2024. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  9. [17]

    Justin D Harris and Bo Waggoner. 2019. Decentralized & Collaborative AI on Blockchain. arXiv preprint arXiv:1907.07247 (2019)

    2019 arXiv

  10. [18]

    Cheng-Ping Hsieh, Simengs Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, and Boris Ginsburg. 2024. RULER: What’s the Real Context Size of Your Long-Context Language Models? arXiv preprint arXiv:2404.06654 (2024)

    2024 arXiv

  11. [19]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations

    2022

  12. [20]

    Max Jaderberg, Valentin Dalibard, Simon Osindero, Wojciech M Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Simonyan, Chrisantha Fernando, and Koray Kavukcuoglu. 2017. Population Based Training of Neural Networks. arXiv preprint arXiv:1711.0...

    2017 arXiv

  13. [21]

    Haifeng Jin, Qingquan Song, and Xia Hu. 2019. Auto-Keras: An Efficient Neural Architecture Search System. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . ACM, 1946–1956

    2019

  14. [22]

    Black Forest Labs. 2024. FLUX.1: A 12 billion parameter rectified flow transformer for text-to-image synthesis. https://huggingface.co/black-forest-labs/FLUX.1- dev

    2024

  15. [23]

    Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. 2017. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization. Journal of Machine Learning Research 18, 1 (2017), 6765–6816

    2017

  16. [24]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (Proceedings of ...

    2017

  17. [25]

    Meta. 2024. Llama 3.3 70B Instruct. https://huggingface.co/meta-llama/Llama- 3.3-70B-Instruct

    2024

  18. [26]

    Benny Moldovanu and Aner Sela. 2001. The Optimal Allocation of Prizes in Contests. American Economic Review 91, 3 (2001), 542–558

    2001

  19. [27]

    Opentensor Foundation. 2025. Dynamic TAO (dTAO) Proposal and Implementa- tion. Deployed February 13, 2025. https://docs.bittensor.com/dynamic-tao/dtao- faq

    2025

  20. [28]

    WMT Organizers. 2019. ACL 2019 Fourth Conference on Machine Translation (WMT19), Shared Task: Machine Translation of News . http://www.statmt.org/ wmt19/translation-task.html

    2019

  21. [29]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems 35 ...

    2022

  22. [30]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving La- tent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  23. [31]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2024. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. arXiv preprint arXiv:2305.18290 (2024)

    2024 arXiv

  24. [32]

    Yuma Rao, Jacob Steeves, Ala Shaabana, Daniel Attevelt, and Matthew McA- teer. 2021. BitTensor: A Peer-to-Peer Intelligence Market. arXiv preprint arXiv:2003.03917 (2021). https://arxiv.org/abs/2003.03917v2

    2021 arXiv

  25. [33]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 22500–22510

    2023

  26. [34]

    Mahdi Shayan, Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. 2021. Biscotti: A Ledger for Private and Secure Peer-to-Peer Machine Learning. arXiv preprint arXiv:1811.09904 (2021)

    2021 arXiv

  27. [35]

    Yixuan Tang and Yi Yang. 2024. MultiHop-RAG: Benchmarking Retrieval- Augmented Generation for Multi-Hop Queries. arXiv preprint arXiv:2401.15391 (2024)

    2024 arXiv

  28. [36]

    Open Thoughts Team. 2025. OpenThoughts-114k: Open synthetic reasoning dataset with 114k high-quality examples. https://huggingface.co/datasets/open- thoughts/OpenThoughts-114k

    2025

  29. [37]

    Abhishek Thakur. 2024. AutoTrain: No-code training for state-of-the-art models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Computational Linguistics, 419–423. https://aclanthology.org/2024.e...

    2024

  30. [38]

    Yimeng Wei, Yingjie Zhu, Yunzhuo Zheng, Jingwei Chai, and Jingjing Chen. 2024. Exploring the Use of Abusive Generative AI Models on Civitai. arXiv preprint arXiv:2407.12876 (2024)

    2024 arXiv

  31. [39]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 Technical Report. arXiv preprint arXiv:2407.10671 (2024)

    2024 arXiv

  32. [40]

    Binghai Zhang, Zhongtao Liu, Colin Cherry, and Orhan Firat. 2024. When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method. arXiv preprint arXiv:2402.17193 (2024)

    2024 arXiv

  33. [41]

    Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020. Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation. arXiv preprint arXiv:2004.11867 (2020)

    2020 arXiv

  34. [42]

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. AdaLoRA: Adap- tive Budget Allocation for Parameter-Efficient Fine-Tuning. In International Conference on Learning Representations

    2023

  35. [2023]

    arXiv preprint arXiv:2306.01436 (2023)

    Multi-Objective Population Based Training. arXiv preprint arXiv:2306.01436 (2023)

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.