CoChat
Sign Up
Cited by 17Topic ModelingDomain Adaptation and Few-Shot LearningNatural Language Processing Techniques

Laboratory-Scale AI: Open-Weight Models are Competitive with ChatGPT Even in Low-Resource Settings

Open access
Wolfe, Robert, Slaughter, Isaac, Han, Bin, Wen, Bingbing, Yang, Yiwei, Rosenblatt, Lucas, Herman, Bernease, Brown, Eva +3 more·2024·The 2024 ACM Conference on Fairness, Accountability, and Transparency
DOI 10.1145/3630106.3658966
Open this paper in CoChat — free account

Abstract

The rapid proliferation of generative AI has raised questions about the competitiveness of lower-parameter, locally tunable, open-weight models relative to high-parameter, API-guarded, closed-weight models in terms of performance, domain adaptation, cost, and generalization. Centering under-resourced yet risk-intolerant settings in government, research, and healthcare, we see for-profit closed-weight models as incompatible with requirements for transparency, privacy, adaptability, and standards of evidence. Yet the performance penalty in using open-weight models, especially in low-data and low-resource settings, is unclear.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper evaluates whether small open-weight 7B instruction-tuned models can replace GPT-4-Turbo in low-resource, domain-specific settings using a single low-cost GPU. It finds that after qLoRA fine-tuning, open models are competitive or better on several tasks at lower cost, while showing mixed results on bias, privacy, and abstention.

Key findings

  • Fine-tuned open models can match or exceed GPT-4-Turbo on domain tasks: Mistral-7B-Instruct reaches .75 accuracy on climate fact-checking versus .72 for three-shot GPT-4-Turbo; fine-tuned LLaMA-2-7B-Chat reaches .74 fact-checking accuracy and .97 entity-resolution accuracy/F1, while GPT-4-Turbo reaches .97 entity-resolution accuracy/F1 in few-shot settings.
  • Open models are cheaper for the climate fact-checking task: fine-tuned Mistral-7B-Instruct inference costs $.31 versus $2.65 for zero-shot GPT-4-Turbo, and fine-tuning an open model costs $3.18–$3.49 versus $3.86 for one-shot GPT-4-Turbo inference.
  • Modest data suffices for adaptation: LLaMA-2-Chat-7B reaches BERTScore-F1 of .79 on clinical dialogue summarization after 20% of training samples (240 samples), within .02 of its best, and .97 F1 on entity resolution after 40%; Mistral-7B-Instruct reaches .71 fact-checking accuracy after 2,000 samples.
  • Responsible-use results are mixed: privately fine-tuned Mistral-7B-Instruct approaches non-private performance at ε=20 (.59 accuracy vs .65 non-private), fine-tuned LLaMA-2-7B-Chat achieves abstention score .99 versus .53 for fine-tuned GPT-3.5-Turbo, but closed models still outperform open models on toxicity bias and fine-tuning did not increase worst-group accuracy for any open model.

Methods

  • Compared three open instruction-tuned 7B models (Mistral-7B-Instruct-v0.1, Falcon-7B-Instruct, LLaMA-2-Chat-7B) against GPT-3.5-Turbo and GPT-4-Turbo in zero-shot, few-shot, and fine-tuned regimes; open models were fine-tuned for a single dataset epoch, while GPT-4-Turbo could not be fine-tuned.
  • Core tasks: entity resolution (custom public-records dataset; 700 train, 100 validation, 200 test), climate fact-checking (Climate-FEVER; 4,298 train, 1,842 validation, 1,535 test), and clinical dialogue summarization (MTS-Dialog; 1,201 train, 100 validation, 400 test); additional responsible-use tasks included MedQA-TF for differential privacy, CivilComments-WILDS toxicity subset (800 train, 100 validation, 200 test), and QASPER abstention (2,593 train, 1,005 validation, 1,451 test).
  • Used 4-bit quantization and qLoRA (rank 32, dropout .05, batch size 1) on a $0.32/hour GCP spot instance with a 16GB Nvidia T4 GPU; metrics included accuracy, F1, BLEU, BERTScore F1, worst-group accuracy, and abstention rate; cost computed from OpenAI token pricing and logged cloud runtime.

Limitations

  • Stated: open models still depend on expensive, resource-intensive, sometimes opaque pretraining that cannot be circumvented; closed-model results may not be reproducible if OpenAI changes or removes API models; carbon cost of closed models could not be reliably modeled.
  • Stated: the authors had difficulty obtaining cloud GPU instances, quantized open models ran much more slowly than closed models, small batch sizes impede differentially private fine-tuning, open models lag on bias mitigation, and fine-tuning has mixed effects on abstention.
  • Evident from the text: GPT-4-Turbo could not be fine-tuned, so fine-tuned comparisons exclude it; the tasks are described as proxies to real-world applications and may not be desirable without human supervision; the cost analysis omits persistent disk storage, static IPs, more reliable cloud instances, and output tokens, which the authors estimate at less than 1% of inference cost.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3630106.3658966

Cites

63 of 63 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

16 of 22 cited sources checked · 9 supported · 0 flagged · accuracy not yet measured

  • Premisecitation
    § 1 INTRODUCTION

    Eight-bit and four-bit quantization allow very large models to run on affordable commercial-grade GPUs.

    “Eight-bit and four-bit quantization allow very large models to run on affordable commercial-grade GPUs [20, 22].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [20] [20] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Llm. int8 (): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339 , (2022), . - source
    • cites [22] [22] Tim Dettmers and Luke Zettlemoyer. 2023. The case for 4-bit precision: k-bit inference scaling laws. In International Conference on Machine Learning . PMLR, , , 7750-7774. - source

      The Abstract (p. 1) says 8-bit and 4-bit quantization reduce model memory footprints, and Recommendations & Future Work (p. 7) gives a concrete example of GPU memory determining which large quantized models fit. The paper does not establish that the GPUs needed to run those models are affordable or commercial-grade. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation
    § 1 INTRODUCTION

    Details of the architecture, parameter counts, and training data of GPT-3.5-Turbo and GPT-4-Turbo are omitted or glancingly described in OpenAI's technical reports.

    “details of the architecture, parameter counts, and training data of GPT-3.5-Turbo and GPT-4-Turbo are omitted or glancingly described in the company's technical reports [41].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [41] [41] OpenAI, :, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fo source

      Section 2, “Scope and Limitations of this Technical Report” (p. 2), says the GPT-4 report gives no further architectural details, including model size, and only broadly describes its training data. That supports the point for GPT-4, but this paper does not establish the same point for the specifically named GPT-3.5-Turbo or GPT-4-Turbo variants. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Premisecitation
    § 3.3 Cloud Infrastructure

    Fine-tuning in four-bit precision is competitive with fine-tuning in higher precision.

    “because fine-tuning in four-bit precision is competitive with fine-tuning in higher precision [21]”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [21] [21] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314 , (2023), . - source
  • Premisecitation
    § 5.1 Differential Privacy - Privately Fine-tuned Open Models Approach Non-Private Performance

    Recent work demonstrates the potential to train general purpose models using differentially private fine-tuning on sensitive data.

    “While challenging in the context of language models [50], recent work [70, 73] demonstrates the potential to train general purpose models using differentially private fine-tuning on sensitive data [72].”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [70] [70] Lukas Wutschitz, Huseyin A. Inan, and Andre Manoel. 2022. dp-transformers: Training transformer models with differential privacy. https: //www.microsoft.com/en-us/research/project/dp-transformers. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [73] [73] Da Yu, Saurabh Naik, Arturs Backurs, Sivakanth Gopi, Huseyin A Inan, Gautam Kamath, Janardhan Kulkarni, Yin Tat Lee, Andre Manoel, Lukas Wutschitz, et al. 2021. Differentially private fine-tuning of language models. arXiv preprint arXiv:2110.06500 , (2021), . - source
    • cites [72] [72] Ashkan Yousefpour, Igor Shilov, Alexandre Sablayrolles, Davide Testuggine, Karthik Prasad, Mani Malek, John Nguyen, Sayan Ghosh, Akash Bharadwaj, Jessica Zhao, et al. 2021. Opacus: User-friendly differential privacy library in PyTorch. arXiv preprint arXiv:2109.12298 , (2021), . - source
  • Premisecitation
    § 6.2 Affordances and Challenges of Open Models.

    LLaMA-2-Chat-7B undergoes RLHF.

    “we note that LLaMA-2-Chat-7B also undergoes RLHF [60]”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [60] [60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 , (2023), . - source
  • Classified as probably not a substantial claim. 24
  • May combine two claims. 26
  • The quote only partly states this claim. 1

  • Classified as probably not a substantial claim.

    Premisecitation
    § 1 INTRODUCTION

    Quantized low-rank adaptation (qLoRA) allows large models to be customized to a domain by adding and tuning a modest number of parameters while allowing most pretrained weights to remain fixed.

    “Quantized low-rank adaptation (qLoRA) [21, 32] allow large models to be customized to a domain by adding and tuning a modest number of parameters while allowing most pretrained weights to remain fixed.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [21] [21] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314 , (2023), . - source
    • cites [32] [32] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 , (2021), . - source

      The Abstract (p. 1) and Introduction (pp. 1–2) say that LoRA adapts models to tasks or domains by training small, added low-rank matrices while freezing pretrained weights. The cited paper describes LoRA, not quantized LoRA; the quantization component of “qLoRA” is not supported by our copy. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 3.1 Models

    GPT-4-Turbo holds the zero-shot state-of-the-art on numerous NLP tasks as of this writing, and achieves first place in human evaluations of chat-based models in Chatbot Arena.

    “GPT-4-Turbo holds the zero-shot state-of-the-art on numerous NLP tasks as of this writing, and achieves first place in human evaluations of chat-based models in Chatbot Arena [77].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [77] [77] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv:2306.05685 [cs.CL] source

      The GPT-4 Judgment section describes GPT-4 as a state-of-the-art LLM and Chatbot Arena as a platform where people rate chatbot responses. Section 4.3 reports that GPT-4 outperformed other models on MT-bench. The copy does not identify GPT-4-Turbo, establish zero-shot state-of-the-art performance across numerous NLP tasks, or report a first-place Chatbot Arena ranking for it. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation
    § 1 INTRODUCTION

    For many scientists and public interest practitioners, lack of transparency is a reason to avoid closed models in their work.

    “For many scientists and public interest practitioners, this lack of transparency is at best concerning, and often a reason to avoid such models in their work altogether [38, 46].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [38] [38] Andreas Liesenfeld, Alianda Lopez, and Mark Dingemanse. 2023. Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators. In Proceedings of the 5th international conference on conversational user interfaces . , , 1-6. - source

      The Introduction says proprietary systems lack transparent ethical oversight and that working scientists call for avoiding proprietary models. Section 1.1 explains why openness matters for research. The paper does not establish that many scientists or public-interest practitioners avoid these models in their work, or how often they do so. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [46] [46] Alexis Palmer, Noah A Smith, and Arthur Spirling. 2023. Using proprietary language models in academic research requires explicit justification. Nature Computational Science , (2023), 1-2. - source

      The opening section identifies auditability and reproducibility concerns with closed models, and Additional information says scientists ideally would not use them. The paper does not address public-interest practitioners or establish how many people in either group avoid closed models for that reason. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation
    § 5.3 Abstention - Fine-tuned Open Models Largely Abstain from Emitting Misinformation

    If the model has the necessary information in neither its parametric knowledge nor the user-provided context, the model should abstain from answering to avoid misinforming a user.

    “If the model has the necessary information in neither its parametric knowledge nor the user-provided context, the model should abstain from answering to avoid misinforming a user [68].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [68] [68] Bingbing Wen, Bill Howe, and Lucy Lu Wang. 2024. Characterizing LLM Abstention Behavior in Science QA with Context Perturbations. arXiv preprint arXiv:2404.12452 (2024). - source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation
    § 6.2 Affordances and Challenges of Open Models.

    The HuggingFace ecosystem supports qLoRA, which made customizing quantized models relatively straightforward.

    “The Huggingface ecosystem also supports qLoRA [21], which made customizing quantized models relatively straightforward.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [21] [21] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. arXiv preprint arXiv:2305.14314 , (2023), . - source

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § 5.3 Abstention - Fine-tuned Open Models Largely Abstain from Emitting Misinformation

    Instruction-tuned language models answer questions based on their parametric knowledge or based on context provided as part of the prompt by the user.

    “Instruction-tuned language models answer questions based on their parametric knowledge [45] or based on context provided as part of the prompt by the user [57].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [45] [45] 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), 27730-27744. - source
    • cites [57] [57] Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng-Xin Yong, Harshit Pandey, Michael McKenna, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Tali Bers, Stella Biderman, Leo G

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingcitation
    § 5.1 Differential Privacy - Privately Fine-tuned Open Models Approach Non-Private Performance

    MedQA-TF proved challenging for open models, which performed much lower than the state-of-the-art.

    “Table 4 illustrates how challenging MedQA-TF proved for open models, which performed much lower than the state-ofthe-art [61].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [61] [61] Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Chuck Lau, Ryutaro Tanno, Ira Ktena, et al. 2023. Towards generalist biomedical AI. arXiv preprint arXiv:2307.14334 , (2023), . - source

      Table 2 (A.5 MultiMedBench Examples; page unavailable) reports performance below the state of the art on MedQA, and Table A.3 (Additional information; page unavailable) says Med-PaLM M fell short of Med-PaLM 2. The paper reports MedQA, not MedQA-TF, and these passages do not establish the claim about open models. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § 3.3 Cloud Infrastructure

    Price was generally consistent on GCP and other providers such as AWS and Lambda Labs, within about $.05 per hour.

    “we found price was generally consistent on GCP and other providers such as AWS and Lambda Labs, within about $.05 per hour.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Most laboratory-scale AI applications will be fault-tolerant during fine-tuning.

  • May combine two claims.

    Findingown data
    § 3.3 Cloud Infrastructure

    GPT-4 outperforms other models in few-shot settings, but open models are competitive after fine-tuning with modest assumptions.

    “GPT-4 outperforms other models in few-shot settings, but open models are competitive after fine-tuning with modest assumptions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION

    GPT-3.5-Turbo and open models are comparable to GPT-4-Turbo or exceed its performance after fine-tuning for a single dataset epoch.

    “but GPT-3.5-Turbo and open models are comparable to GPT-4-Turbo or exceed its performance after fine-tuning for a single dataset epoch”
    • Also stated in: § 3.3 Cloud Infrastructure · hedged
      “GPT-4 outperforms other models in few-shot settings, but open models are competitive after fine-tuning with modest assumptions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.2 Performance - Fine-tuned Open Models Can Outperform Closed Models

    Fine-tuned Mistral-7B-Instruct achieves a higher BLEU score but not a higher BERT score than GPT-3.5-Turbo on medical summarization.

    “With fine-tuning, though, Mistral-7B-Instruct outperforms GPT-4-Turbo few shot, achieving higher BLEU score (but not higher BERT score) than GPT-3.5-Turbo, while fine-tuned LLaMA-2-7B-Chat and Falcon-7B-Instruct achieve results competitive with few-shot GPT-4-Turbo.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION

    Fine-tuned open models exhibit performance comparable to general-purpose base chat models on tasks not reflected in the model's fine-tuning dataset, and in some cases exceed it.

    “We find that fi ne-tuned open models exhibit performance comparable to general-purpose base chat models , and in some cases exceed it”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.2 Performance - Fine-tuned Open Models Can Outperform Closed Models

    Fine-tuning for a single dataset epoch yields open models that are competitive and in some cases even outperform GPT-4-Turbo and fine-tuned GPT-3.5-Turbo.

    “Fine-tuning for a single dataset epoch, however, yields open models that are competitive and in some cases even outperform GPT-4-Turbo and fine-tuned GPT-3.5-Turbo.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION

    After 2,000 fact-checking samples, Mistral-7B-Instruct achieves accuracy of .71, comparable to fine-tuned GPT-3.5-Turbo.

    “After 2,000 fact-checking samples, Mistral-7B-Instruct achieves accuracy of .71, comparable to fine-tuned GPT-3.5-Turbo.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION

    Open models exhibit moderate bias that fine-tuning largely fails to mitigate.

    “that open models exhibit moderate bias that fine-tuning largely fails to mitigate”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION

    After 20% of the fine-tuning dataset (240 samples), the clinical dialogue summarization model achieves accuracy of .79 on the test dataset, within .02 of its best.

    “After 20% of the fine-tuning dataset (240 samples), the summarization model achieves accuracy of .79 on the test dataset, within .02 of its best.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning
    § 6.3 Limitations and Future Work.

    The authors could not reliably model the carbon cost of closed models due to uncertainties about the exact hardware used to run these models, the location of the data centers on which they run, and practices such as batching user inputs.

    “Finally, we could not reliably model the carbon cost of closed models due to uncertainties about the exact hardware used to run these models, the location of the data centers on which they run, and practices such as batching user inputs, which may allow for economies of scale.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.3 Abstention - Fine-tuned Open Models Largely Abstain from Emitting Misinformation

    Fine-tuning improves Mistral-7B-Instruct, Falcon-7B-Instruct and LLaMA-2-7B-Chat on QASPER, but performance does not approach GPT-4-Turbo.

    “Fine-tuning improves Mistral-7B-Instruct, Falcon-7B-Instruct and LLaMA-2-7B-Chat, but performance does not approach GPT-4-Turbo on this challenging task.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingown data
    § 5.3 Abstention - Fine-tuned Open Models Largely Abstain from Emitting Misinformation

    GPT4-Turbo excels in few-shot settings on QASPER.

    “GPT4-Turbo excels in few-shot settings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning
    § 6.2 Affordances and Challenges of Open Models.

    The authors experienced consistent difficulties obtaining results due to a lack of available cloud resources.

    “We experienced consistent difficulties obtaining results due to lack of available cloud resources.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.2 Toxicity Bias - Open Models Improve with Fine-Tuning, But Lag Behind Closed Models

    Fine-tuning marginally improves toxicity classification accuracy in open models, but closed models still consistently outperform them.

    “Table 5. Fine-tuning marginally improves toxicity classification accuracy in open models, but closed models still consistently outperform them.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given
    § 4.3 Cost Analysis - Open Models Are More Affordable

    Fine-tuned GPT-3.5-Turbo is the fastest option and ten times faster than open models.

    “Closed models excel on runtime. Fine-tuned GPT-3.5-Turbo is the fastest option, and ten times faster than open models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Cost Analysis - Open Models Are More Affordable

    Few-shot GPT-4-Turbo requires 1.5 times as long as few-shot GPT-3.5-Turbo, but is three times as fast as open models.

    “Few-shot GPT-4-Turbo requires 1.5 times as long as few-shot GPT-3.5-Turbo, but is three times as fast as open models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given
    § 6.2 Affordances and Challenges of Open Models.

    The Supervised Fine-Tuning trainer class provided by the TRL library made adapting open language models relatively simple, and primarily dependent on the organization of the data.

    “The Supervised Fine-Tuning trainer class provided by the TRL library made adapting open language models relatively simple, and primarily dependent on the organization of our data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.4 Data Responsiveness - Modest Fine-tuning Can Make Open Models Competitive

    Fine-tuning improvements emerge during the first 50% of the training data, only a few hundred training samples in the case of Medical Summarization and Entity Resolution.

    “Left: Fine-tuning improvements emerge during the first 50% of the training data, only a few hundred training samples in the case of Medical Summarization and Entity Resolution.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § 6.2 Affordances and Challenges of Open Models.

    The most intractable problem the authors encountered in fine-tuning their own models was the difficulty of obtaining cloud instances equipped with even low-cost GPU hardware.

    “The most intractable problem we encountered in fine-tuning our own models lay in the difficulty of obtaining cloud instances equipped with even low-cost GPU hardware.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.2 Toxicity Bias - Open Models Improve with Fine-Tuning, But Lag Behind Closed Models

    Fine-tuned GPT-3.5-Turbo matches three-shot GPT-4-Turbo on overall accuracy, but not Worst-Group accuracy.

    “Fine-tuned GPT-3.5-Turbo matches three-shot GPT-4-Turbo on overall accuracy, but not Worst-Group accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 5.2 Toxicity Bias - Open Models Improve with Fine-Tuning, But Lag Behind Closed Models

    The strongest performing model of the group was three-shot GPT-4-Turbo, which exceeded other models in both Mean and Worst-Group accuracy.

    “The strongest performing model of the group was three-shot GPT-4-Turbo, which exceeded other models in both Mean and Worst-Group accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § 1 INTRODUCTION

    While open models exhibit greater bias, they offer greater privacy affordances than closed models, and in some cases abstain more reliably after fine-tuning.

    “While open models exhibit greater bias, they offer greater privacy affordances than closed models, and in some cases abstain more reliably after fine-tuning.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data
    § 3.3 Cloud Infrastructure

    The study used a consistent cloud environment, allowing comparison of the cost and runtime of open vs. closed models.

    “We use a consistent cloud environment, allowing comparison of the cost and runtime of open vs. closed models.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Most laboratory-scale AI applications will be fault-tolerant during fine-tuning.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 3.1 Models

    OpenAI has not disclosed parameter counts for GPT-3.5-Turbo and GPT-4-Turbo, but studies suggest they are much larger than open models.

    “OpenAI has not disclosed parameter counts for GPT-3.5Turbo and GPT-4-Turbo, but studies suggest they are much larger than open models [48].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] [48] Dylan Patel and Gerald Wong. 2023. GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE. https://www.semianalysis.com/p/gpt4-architecture-infrastructure. [Accessed 19-01-2024]. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning
    § 3.2 Model Evaluation

    GPT-4-Turbo could not be fine-tuned because fine-tuning was available only via an experimental program.

    “We are unable to fine-tune GPT-4-Turbo, for which fine-tuning is available only via an experimental program.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 7 CONCLUSION

    Small open models are responsive to user data.

    “responsive to user data”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 5.1 Differential Privacy - Privately Fine-tuned Open Models Approach Non-Private Performance

    Differentially private deep learning has been adopted to protect users and avoid legal risks of sensitive data use.

    “Differentially private (DP) deep learning (using a privatized gradient descent optimizer [1]) has been adopted to protect users and avoid legal risks of sensitive data use [27, 29].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [27] [27] Badih Ghazi, Noah Golowich, Ravi Kumar, Pasin Manurangsi, and Chiyuan Zhang. 2021. Deep learning with label differential privacy. Advances in neural information processing systems 34 (2021), 27131-27145. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [29] [29] Maoguo Gong, Yu Xie, Ke Pan, Kaiyuan Feng, and Alex Kai Qin. 2020. A survey on differentially private machine learning. IEEE computational intelligence magazine 15, 2 (2020), 49-64. - source

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 7 CONCLUSION

    Small open models are robust to fine-tuning.

    “robust to fine-tuning”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 4.3 Cost Analysis - Open Models Are More Affordable

    Fine-tuned open models are much less expensive than GPT-4-Turbo and more performant than few-shot closed models.

    “Fine-tuned open models are much less expensive than GPT-4-Turbo, and more performant than few-shot closed models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning
    § 7 CONCLUSION

    Laboratory-scale AI can serve as a basis for future scientific and public interest work, enabling practitioners to customize models without needing to rely on closed, API-based AI.

    “We contend that laboratory-scale AI can serve as a basis for future scientific and public interest work, enabling practitioners to customize models without needing to rely on closed, API-based AI.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 6.3 Limitations and Future Work.

    Results from closed models may not be reproducible if OpenAI changes or removes models from its API without notifying the end user.

    “We also acknowledge that results from closed models may not be reproducible, should OpenAI change or remove models from its API, potentially without notifying the end user.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § 5.1 Differential Privacy - Privately Fine-tuned Open Models Approach Non-Private Performance

    A challenge of noisy, privatized updates is that batch size needs to be large, posing issues for lab-scale approaches that use smaller batches.

    “A challenge of noisy, privatized updates is that batch size needs to be large, posing issues for lab-scale approaches that use smaller batches.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § 7 CONCLUSION

    Small open models are cost-efficient.

    “they are cost-efficient”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 6.1 The Viability and Implications of Laboratory-Scale AI.

    The authors' work provides empirical support for the viability of adopting a laboratory-scale approach to AI that prioritizes user autonomy, privacy, fairness, and transparency while maintaining much of the performance and usability offered by industry-dominant corporate models.

    “Our work provides empirical support for the viability of adopting a 'laboratory-scale' approach to AI that prioritizes user autonomy, privacy, fairness, and transparency while maintaining much of the performance and usability offered by industry-dominant corporate models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § 6.1 The Viability and Implications of Laboratory-Scale AI.

    Laboratory-scale AI presents an option for scientists and public interest technologists who have good reason to avoid closed models that cannot be accessed except via a call to an API.

    “Laboratory-scale AI centers the domain-specific, responsible application of small, open models, presenting an option for scientists and public interest technologists who have good reason to avoid closed models that cannot be accessed except via a call to an API.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Closed models cannot be accessed except via a call to an API.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § 6.2 Affordances and Challenges of Open Models.

    The slower inference speed of quantized open models is related in part to the choice of a low-end GPU.

    “This is related in part to our choice of a low-end GPU”
    Quote in the paperNot a substantial claim
43 of 43

  • The authors state this as an assumption.

    Closed-weight models are incompatible with requirements for transparency, privacy, adaptability, and standards of evidence in under-resourced, risk-intolerant settings.

    “Centering under-resourced yet risk-intolerant settings in government, research, and healthcare, we see for-profit closed-weight models as incompatible with requirements for transparency, privacy, adaptability, and standards of evidence.”

  • The paper relies on this without support.

    The performance penalty in using open-weight models, especially in low-data and low-resource settings, is unclear.

    “Yet the performance penalty in using open-weight models, especially in low-data and low-resource settings, is unclear.”

  • The authors state this as an assumption.

    Most laboratory-scale AI applications will be fault-tolerant during fine-tuning.

    “Because we expect that most laboratory-scale AI applications will be fault-tolerant during fine-tuning, we use spot instances”

  • The authors state this as an assumption.

    The selected tasks are proxies to real-world applications.

    “We acknowledge it may not be desirable to use an LM in a setting like clinical dialogue summarization or fact-checking, especially without human supervision, and that our tasks are proxies to real-world applications.”

  • The paper relies on this without support.

    Closed models have a process used to mitigate bias and prevent the closed model from generating harmful or inaccurate output.

    “One of the presumed advantages offered by closed models is the process used to mitigate bias and prevent the closed model from generating harmful or inaccurate output.”

  • The paper relies on this without support.

    Open models have undergone expensive, resource intensive pretraining on large-scale, sometimes opaque datasets.

    “we acknowledge that open models have undergone expensive, resource intensive pretraining on large-scale, sometimes opaque datasets.”

  • The paper relies on this without support.

    qLoRA cannot equip users with a means of circumventing pretraining.

    “While libraries like qLoRA help to enable adaptations of pretrained models, they cannot equip us with a means of circumventing pretraining”

  • The paper relies on this without support.

    Pretraining remains the only reliable means of producing a fluent, general-purpose base model.

    “which at this time remains the only reliable means of producing a fluent, general-purpose base model.”

  • The paper relies on this without support.

    Closed models cannot be accessed except via a call to an API.

    “closed models that cannot be accessed except via a call to an API”
9 of 9

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • GPT-4-Turbo zero-shot cost differs from Table 3

    Minor

    The introduction reports $2.65 for zero-shot GPT-4-Turbo inference on the climate fact-checking test set, while Table 3 reports $2.60.

  • Summarization BERTScore-F1 is called accuracy

    Minor

    The introduction calls the .79 clinical summarization result accuracy, but the results section identifies that same 20%-of-training result as BERTScore-F1.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 11

  1. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901.
  2. Andrew M Dai and Quoc V Le. 2015. Semi-supervised sequence learning. Advances in neural information processing systems 28 (2015) .
  3. 10.18653/v1 · DOI
  4. Maoguo Gong, Yu Xie, Ke Pan, Kaiyuan Feng, and Alex Kai Qin. 2020. A survey on differentially private machine learning. IEEE computational intelligence magazine 15, 2 (2020), 49–64.
  5. Lefteris Loukas Ilias Stogiannidis Prodromos Malakasiotis and Stavros Vassos. 2023. Breaking the Bank with ChatGPT: Few-Shot Text Classification for Finance. In Proceedings of the Fifth Workshop on Financial Technology and Natural Language Processing and the Second Multimodal AI For Financial Forecasting Chung-Chi Chen Hiroya Takamura Puneet Mathur Remit Sawhney Hen-Hsen Huang and Hsin-Hsi Chen (Eds.). - Macao 74–80. https://aclanthology.org/2023.finnlp-1.7
  6. OpenAI. 2022. Introducing ChatGPT. OpenAI Blog, (Nov 2022),.
  7. Jon Porter. 2023. ChatGPT continues to be one of the fastest-growing services ever. The Verge, (Nov 2023),.
  8. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9.
  9. Victor Sanh Albert Webson Colin Raffel Stephen H. Bach Lintang Sutawika Zaid Alyafeai Antoine Chaffin Arnaud Stiegler Teven Le Scao Arun Raja Manan Dey M Saiful Bari Canwen Xu Urmish Thakker Shanya Sharma Eliza Szczechla Taewoon Kim Gunjan Chhablani Nihal V. Nayak Debajyoti Datta Jonathan Chang Mike Tian-Jian Jiang Han Wang Matteo Manica Sheng Shen Zheng-Xin Yong Harshit Pandey Michael McKenna Rachel Bawden Thomas Wang Trishala Neeraj Jos Rozen Abheesht Sharma Andrea Santilli Thibault Fevry Jason Alan Fries Ryan Teehan Tali Bers Stella Biderman Leo Gao Thomas Wolf and Alexander M. Rush. 2022. Multitask Prompted Training Enables Zero-Shot Task Generalization. In International Conference on Learning Representations. . https://research.ibm.com/publications/multitask-prompt-tuning-enables-zero-shot-task-generalization
  10. Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Llion Jones Aidan N Gomez Łukasz Kaiser and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017) .
  11. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. BERTScore: Evaluating Text Generation with BERT. In International Conference on Learning Representations.,,.

Cited by

Sorted by citations

Related

Authors
Wolfe, R., Slaughter, I., Han, B., Wen, B., Yang, Y., Rosenblatt, L., Herman, B., Brown, E., Qu, Z., Weber, N., & Howe, B.
Affiliations
  • University of Washington (US)
  • New York University (US)
  • University of Washington, United States
ORCID
Venue
The 2024 ACM Conference on Fairness, Accountability, and Transparency
Publisher
ACM
Published
2024-06-03
Pages
1199-1210
DOI
10.1145/3630106.3658966
OpenAlex
W4399116034
Semantic Scholar
a17144858b060c24a46aebbef969218737307bf0
Cited by
17
References
29
Citations by year
2026: 10 · 2025: 7
FWCI
5.50
Type
conference paper
Language
en
Open access
Licence
Locations
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-sa)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Keywords
open-weight models, resource-limited settings, small language models, ChatGPT comparison, domain adaptation, model generalization, transparency, privacy

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review