UTSA-NLP at ArchEHR-QA 2025: Improving EHR Question Answering via Self-Consistency Prompting

Anthony Rios; David Perry; Joshua Gardner; Sara Shields-Menard; Zach Reimers

REVIEW 3 major objections 5 minor 23 references

UTSA-NLP at ArchEHR-QA 2025: Improving EHR Question Answering via Self-Consistency Prompting

T0 review · 3 major / 5 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 In EHR question answering, finding the right sentences matters more than generating the answer: a small 8B selector plus a 70B generator beats the shared-task baseline, and perfect sentence selection almost doubles the score.

desk verdict A competent shared-task system paper; the oracle headroom makes the sentence-selection point convincingly, but the thresholding evidence is partly circular and the table/prose numbers are swapped. read the letter →

arxiv 2506.05589 v1 pith:X2LXPNNG submitted 2025-06-05 cs.CL

Sara Shields-Menard , Zach Reimers , Joshua Gardner , David Perry , Anthony Rios This is my paper

classification cs.CL

keywords clinicalquestionansweringelectronichealthrecordsself-consistencythresholdingsentenceselectionfew-shotpromptinglargelanguagemodelsgrounded

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

This paper argues that the main limit on grounded question answering over electronic health records is not answer writing but evidence selection: deciding which sentences in a patient's record actually address the clinician's question. The authors build a two-stage pipeline in which a few-shot prompt labels each sentence essential, supplementary, or not relevant, and a second model composes a citation-backed answer of at most 75 words from the selected sentences. Their best configuration, an 8B sentence selector paired with a 70B generator, scores 42.4 on the development set against a shared-task baseline of 35.9. An oracle with ground-truth sentence labels reaches 74.6, and ablations that remove the self-consistency thresholding step drop performance to the mid-to-high 20s, supporting the claim that reliable sentence selection is what moves the overall score.

What carries the argument

The mechanism is self-consistency decoding with thresholding. The same few-shot prompt independently classifies each EHR sentence; the label is sampled 20 times, and majority counting is replaced by low thresholds (2/20 for essential, 1/20 for supplementary) that bias the classifier against the dominant not-relevant class. This step is what the ablations identify as the main source of the gain over the baseline, and it feeds a separate zero-shot generation stage that assembles the final 75-word response with citations.

What would settle it

On the 100-case test set, run the best configuration three ways—with the manual thresholds, without thresholding, and with thresholds tuned on a different split—and compare overall scores; the claim that thresholding is central fails if the thresholded variant does not reproducibly beat the unthresholded one or if alternative thresholds perform equally well.

Watch

Extended reading notes

Core claim

The central discovery is that accurate sentence selection is the binding constraint in EHR question answering. With ground-truth sentence labels, the system's overall score is 74.58, against 42.4 for the best automatic configuration, meaning most of the remaining error is in choosing sentences rather than composing answers. The paper also finds that a smaller LLaMA 3.1 8B model labels sentences more reliably than a quantized 70B model, and that a simple self-consistency rule—sampling the label 20 times at temperature 1.0 and classifying a sentence as essential when at least 2 of 20 samples say so, supplementary when at least 1 does, and not relevant otherwise—produces the best selection and the best final answers.

Load-bearing premise

The thresholds, few-shot examples, and model choices were tuned on a 20-case development set, so the paper's measured gains assume those 20 cases are representative enough of the 100-case test set for the manual tuning to transfer.

Editorial extensions

If this is right

Improving the sentence selector should yield larger gains than scaling the answer generator, since the oracle gap (74.6 vs 42.4) is almost entirely on the selection side.
Deploying a small selector with a large generator is a viable cost trade-off: the 8B selector beat the 70B selector, so the expensive model can be reserved for final composition.
Thresholding offers a no-training calibration lever for rare-but-critical classes, shifting the classifier toward recall of essential sentences.
The supplementary class is the main error source; merging it into a binary relevant/not-relevant decision, or better modeling of it, directly targets the largest remaining gap.
Overall response quality can be expected to track sentence-selection F1, so progress on the selector should translate point-for-point into better grounded answers.

Reading between the lines

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

The 2/20 and 1/20 thresholds effectively encode a prior over how often relevant sentences appear; a data-driven version of this prior could be fit per note type or per hospital and might transfer better than the manual values.
The 8B-over-70B result may reflect quantization or prompt sensitivity rather than a general law; a controlled comparison with equally quantized models would test that.
The error analysis shows the selector leans on lexical overlap (e.g., matching an acronym) and misses clinical paraphrases; a scorer built on textual entailment or trained retrieval could attack exactly those failures.
If this selection-first recipe generalizes, other grounded generation tasks—such as answering from patient portals or discharge summaries—could adopt the same small-selector/large-generator split.

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.

Desk Editor's Note

read the letter

Quick take: this is a decent shared-task system paper, not a methodological breakthrough. The most useful thing is the oracle analysis: perfect sentence selection lifts the dev score from 42.4 to 74.58 (strict), which backs up the central claim that retrieval quality, not answer generation, is the binding constraint in these pipelines. That is a clean empirical observation and the main reason to read the paper.

What is genuinely new here is narrow: applying self-consistency with manually fixed thresholds (essential: at least 2 of 20 samples, supplementary: at least 1 of 20) to sentence-level relevance classification in EHRs. The paper also compares an 8B classifier against a quantized 70B one and finds the 8B does better; interesting, but the comparison is confounded by quantization and prompt sensitivity, and the authors acknowledge this in the Limitations.

The soft spots, in rough order of importance:

1. Internal inconsistency: Section 3.2 says removing thresholding alone gives 26.34 and removing both gives 29.31, but Table 1 lists 29.3 and 26.3 respectively. Those numbers are swapped. Easy to fix, but it needs a correction.

2. The thresholding benefit is measured on the same 20-case dev set used to tune the thresholds. The 42.4 vs 29.3 gap from thresholding is not backed by any held-out ablation on the test set; the test results only show the full pipeline beats the baseline. So the abstract's claim that 'self-consistency with thresholding helps' is partly a fit to 20 examples. The authors do flag in Limitations that the thresholds were tuned manually and may not generalize, which is the right caveat, but the abstract states it more confidently.

3. No significance testing anywhere, and the dev set is tiny. Typical for shared tasks, but it means the 8B-vs-70B and thresholding results should be read as suggestive, not definitive.

4. The thresholds themselves are very low (2/20 for essential), which biases toward predicting relevant and explains the high false-positive counts in the confusion matrices. That is a stated trade-off, not a hidden flaw.

Who is this for? People building evidence-selection components for grounded QA, especially in clinical domains, and anyone comparing small vs large LLMs for classification. The oracle headroom result is worth remembering.

My take: it deserves a serious referee. The empirical core, that sentence selection is the bottleneck, is well supported. I would recommend acceptance after minor revision: fix the swapped numbers, reword the abstract to not overclaim the thresholding result, and ideally report a test-set ablation if the organizers released enough data. Not a transformative paper, but a solid, honest system description with one clean insight.

Referee Report

3 major / 5 minor

Summary. This paper describes UTSA-NLP's submission to the ArchEHR-QA 2025 shared task. The system splits the task into sentence-level relevance classification and answer generation: a LLaMA 3.1 8B model labels each EHR sentence as essential, supplementary, or not relevant using few-shot prompting, 20 self-consistency samples, and two manually set thresholds; a LLaMA 3.1 70B model then generates a citation-supported answer from the selected sentences. On the 20-case development set, the best configuration (8B selector, lenient evaluation) reaches 42.4 overall versus 35.9 for the shared-task baseline, while an oracle with gold essential sentences reaches 74.58. The test-set submission scores 40.45 versus the organizers' 30.7 baseline. The paper argues that accurate sentence selection is the main bottleneck and that self-consistency with thresholding improves selection reliability.

Significance. If the claims hold, the paper provides useful evidence for the shared-task community: a compact two-stage prompting pipeline can substantially beat the baseline, and the large oracle gap isolates sentence selection as the bottleneck. The paper's strengths are the explicit oracle comparison, the 8B-versus-70B comparison, the confusion-matrix error analysis, and a candid Limitations section. The main risk is the small (20-case) development set and the fact that the thresholding gain is reported only on the same set used to tune the thresholds; no held-out ablation is presented. For a shared-task system report this is acceptable as a practical engineering result, but the causal phrasing in the abstract overstates the evidence.

major comments (3)

[§3.2, Table 1] The ablation numbers in the text do not match Table 1: §3.2 says removing thresholding alone reduces overall performance to 26.34 and removing both thresholding and self-consistency to 29.31, while Table 1 lists 29.3 and 26.3, respectively. Please correct the text or the table and clarify the direction of the effect; the current inconsistency makes it impossible to attribute the gain to thresholding versus self-consistency.
[§2.2, §3.2, Table 4] The thresholds (at least 2/20 for essential, at least 1/20 for supplementary) were manually tuned on the 20-case development set, and the only ablation showing their benefit is evaluated on that same set (Table 1). The test set results in Table 4 do not include a condition with thresholding disabled, so the abstract's claim that 'self-consistency with thresholding helps make these decisions more reliable' is not supported by held-out evidence. Please either add a test-set ablation or restrict the claim to the development-set observations.
[§3.2] The oracle results show GT-Strict (74.58) outperforming GT-Lenient (61.90), yet the system results show the opposite ordering (8B-LENIENT 42.4 vs. 8B-STRICT 36.1). Since the lenient configuration is the paper's best system, the paper should explain this crossover; otherwise it is unclear whether the lenient gain is a property of the method or an artifact of noisy classification interacting with the 75-word generation constraint.

minor comments (5)

[§2.1, §3.2] Please state explicitly that the 30 few-shot examples in each leave-one-out fold are drawn only from the remaining 19 development cases; if they may include the held-out case, Table 1 would be contaminated.
[§3.3, Table 4] The table label '8B Results' does not identify whether this is the lenient or strict configuration used in the test submission; please align the notation with Section 3.3.
[References] Liu et al. 2023a and 2023b appear to be the same paper with the same title and venue; please merge the two entries or clearly distinguish them.
[Appendix A] The prompt contains the typo 'anonyminity' instead of 'anonymity'; additionally, fix the inconsistent model spelling 'LL AMA3-8B' in Section 3.2.
[§2.2] The description of self-consistency sampling says 20 independent predictions are made at temperature 1.0, but it does not specify what varies across samples; please clarify whether the random seed or the sampled few-shot examples vary.

Circularity Check

1 steps flagged · score 6.0 of 10

Thresholding benefit is measured on the same development set used to tune the thresholds, so the claim that self-consistency with thresholding helps is partly a fitted result.

fitted input called prediction [Section 2.2 (Self-Consistency and Thresholds) and Section 3.2 (Validation Results) / Table 1]
"Thresholds are applied to determine the final label: a sentence is labeled essential if it appears in at least 2 out of 20 predictions; if not, it is labeled supplementary if it appears at least once; otherwise, it is labeled not relevant. ... We repeated several attempts with these median and mean thresholds, but these high thresholds yielded F1 scores with several false negatives for the essential label. ... We report the results on the dev set using leave-one-out cross-validation in Table 1. ... Removing thresholding alone reduced overall performance to 26.34."

The threshold values (2/20 and 1/20) were manually selected by trying alternatives on the same 20-case development set and keeping values that improved dev-set F1, as the repeated attempts described in Section 2.2 show. The sole evidence that thresholding helps is the Table 1 ablation on that same dev set, comparing the thresholded configuration (42.4) with the unthresholded one (29.3). Because the threshold was chosen precisely because it outperformed other thresholds on this dev set, the reported improvement over no thresholding is partially forced by the tuning criterion; it is an in-sample evaluation, not a held-out prediction. The test set (Table 4) contains no thresholding-off ablation, so no out-of-sample evidence isolates the thresholding effect.

full rationale

The paper's central causal claim--that self-consistency with thresholding improves sentence-selection reliability--rests on a development-set ablation (Table 1) in which the thresholds were manually tuned on that same 20-case dev set (Section 2.2). This is the one genuinely circular element: the thresholds were chosen by trying alternatives and keeping values that improved dev F1, so reporting that thresholding beats no-thresholding on the same dev set is partly a restatement of the selection criterion. Pattern 2 (fitted input called prediction) therefore applies. However, the paper also contains independent, non-circular evidence: the ground-truth oracle comparison (GT-LENIENT 61.90, GT-STRICT 74.58) supports the 'sentence selection is critical' claim using gold labels rather than fitted parameters; the test-set results (Table 4) show the final thresholded 8B pipeline outperforming the organizers' baseline (40.45 vs 30.7) on held-out data; and the 8B-vs-70B comparison is an out-of-sample model comparison. There are no self-citations, imported uniqueness theorems, or renamed known results. The circularity is therefore partial and confined to the thresholding-benefit claim, which is why the score is 6 rather than 0-2.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central pipeline rests on manually tuned thresholds, a small dev set, and the shared task's labels and metrics. No novel physical or conceptual entities are postulated. The free parameters are the sampling count, the two thresholds, and the few-shot example count; all were chosen by hand or by dev-set experiments.

free parameters (4)

Essential threshold count = 2 of 20 self-consistency samples (10%)
Manually set in Section 2.2 after dev-set experiments; the paper states thresholds 'were tuned manually' in Limitations.
Supplementary threshold count = 1 of 20 samples (5%)
Same manual tuning process; lower threshold intended to capture supplementary sentences.
Self-consistency sample count = 20
Chosen after observing limited gains beyond 20 samples on the development set (Section 2.2).
Few-shot example count = 30
Balanced random sample of 30 labeled sentences inserted into the prompt (Section 2.1); no tuning analysis shown.

assumptions (3)

ad hoc to paper The 20-case development set is representative of the test distribution, so thresholds tuned on it generalize.
The paper's main empirical claims use thresholds tuned on the dev set and report performance on the same dev set; generalization to the test set is assumed.
domain assumption The ArchEHR-QA gold labels and scoring metrics are consistent and clinically meaningful.
The evaluation uses shared task labels and automatic metrics (ROUGE, BLEU, SARI, BERTScore, AlignScore, MEDCON) without independent validation.
domain assumption Twenty independent samples at temperature 1.0 give a reliable voting distribution for each sentence's relevance.
Self-consistency thresholding treats sample vote counts as estimates of the model's relevance probability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UTSA-NLP at ArchEHR-QA 2025: Improving EHR Question Answering via Self-Consistency Prompting." pith.science (2026). https://pith.science/paper/X2LXPNNG

@misc{pith2026250605589,
  author       = {Pith},
  title        = {Pith review of: UTSA-NLP at ArchEHR-QA 2025: Improving EHR Question Answering via Self-Consistency Prompting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2LXPNNG}},
  note         = {Machine review of arXiv:2506.05589}
}

read the original abstract

We describe our system for the ArchEHR-QA Shared Task on answering clinical questions using electronic health records (EHRs). Our approach uses large language models in two steps: first, to find sentences in the EHR relevant to a clinician's question, and second, to generate a short, citation-supported response based on those sentences. We use few-shot prompting, self-consistency, and thresholding to improve the sentence classification step to decide which sentences are essential. We compare several models and find that a smaller 8B model performs better than a larger 70B model for identifying relevant information. Our results show that accurate sentence selection is critical for generating high-quality responses and that self-consistency with thresholding helps make these decisions more reliable.

Figures

Figures reproduced from arXiv: 2506.05589 by the authors.

**Figure 2.** Overall Method Figure [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 6 canonical work pages

[1]

Margaret Warton, Mary E

Fatema Akbar, Gloria Mark, E. Margaret Warton, Mary E. Reed, Stephanie Prausnitz, Jeffrey A. East, Mark F. Moeller, and Tracy A. Lieu. 2021. https://doi.org/10.1093/jamia/ocaa229 Physicians’ electronic inbox work patterns and factors associated with high inbox work duration . Journal of the American Medical Informatics Association, 28(5):923--930

work page doi:10.1093/jamia/ocaa229 2021
[2]

Ayers, Adam Poliak, Mark Dredze, and et al

John W. Ayers, Adam Poliak, Mark Dredze, and et al. 2023. https://doi.org/10.1001/jamainternmed.2023.1838 Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum . JAMA Internal Medicine, 183(6):589--596. Published by the American Medical Association

arXiv 2023
[3]

John W. et al. Ayers. 2023. Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine, 183(6):589--596

work page 2023
[4]

Brands, S

M. Brands, S. Gouw, M. Beestrum, R. Cronin, K. Fijnvandraat, and S. Badawy. 2022. https://doi.org/10.2196/43086 Patient-centered digital health records and their effects on health outcomes: Systematic review . Journal of Medical Internet Research, 24(12):e43086

work page doi:10.2196/43086 2022
[5]

Callahan

Alexander et al. Callahan. 2021. https://doi.org/10.1056/CAT.21.0224 Using aggregate patient data at the bedside via an on-demand consultation service . NEJM Catalyst Innovations in Care Delivery, 2

doi:10.1056/cat.21.0224 2021
[6]

Carini, L

E. Carini, L. Villani, A. M. Pezzullo, A. Gentili, A. Barbara, W. Ricciardi, and S. Boccia. 2021. https://doi.org/10.2196/26189 The impact of digital patient portals on health outcomes, system efficiency, and patient attitudes: Updated systematic literature review . Journal of Medical Internet Research, 23(9):e26189

doi:10.2196/26189 2021
[7]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024
[8]

Yunxiang Li, Zihan Li, Kai Zhang, Ruilong Dan, Steve Jiang, and You Zhang. 2023. Chatdoctor: A medical chat model fine-tuned on a large language model meta-ai (llama) using medical domain knowledge. Cureus, 15(6)

2023

Show all 23 references

[9]

Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

2004
[11]

Wright, Barron L

Siru Liu, Aileen P. Wright, Barron L. Patterson, Jonathan P. Wanderer, Robert W. Turer, Scott D. Nelson, Allison B. McCoy, Dean F. Sittig, and Adam Wright. 2023 b . https://doi.org/10.1093/jamia/ocad072 Using ai-generated suggestions from chatgpt to optimize clinical decision ...

2023 doi
[12]

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

2002
[13]

Stephen R. et al. Pfohl. 2024. https://doi.org/10.1038/s41591-024-03258-2 A toolbox for surfacing health equity harms and biases in large language models . Nature Medicine

2024 doi
[14]

Karan et al. Singhal. 2023. Large language models encode clinical knowledge. Nature, 620:172--180

2023
[15]

Karan et al. Singhal. 2025. https://doi.org/10.1038/s41591-024-03423-7 Toward expert-level medical question answering with large language models . Nature Medicine, 31:943--950

2025 doi
[16]

Sarvesh Soni and Dina Demner-Fushman. 2025 a . A dataset for addressing patient's information needs related to clinical course of hospitalization. arXiv preprint

2025
[17]

Sarvesh Soni and Dina Demner-Fushman. 2025 b . Overview of the archehr-qa 2025 shared task on grounded question answering from electronic health records. In The 24th Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks, Vienna, Austria. Association for Co...

2025
[18]

Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. 2016. Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4:401--415

2016
[19]

Wen-wai Yim, Yujuan Fu, Asma Ben Abacha, Neal Snider, Thomas Lin, and Meliha Yetisgen. 2023. Aci-bench: a novel ambient clinical intelligence dataset for benchmarking automatic visit note generation. Scientific data, 10(1):586

2023
[20]

Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. Alignscore: Evaluating factual consistency with a unified alignment function. arXiv preprint arXiv:2305.16739

2023 arXiv
[21]

Kaiyan Zhang, Sihang Zeng, Ermo Hua, Ning Ding, Zhang-Ren Chen, Zhiyuan Ma, Haoxin Li, Ganqu Cui, Biqing Qi, Xuekai Zhu, and 1 others. 2024. Ultramedical: Building specialized generalists in biomedicine. Advances in Neural Information Processing Systems, 37:26045--26081

2024
[22]

Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

2019 arXiv
[23]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[24]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

Margaret Warton, Mary E

Fatema Akbar, Gloria Mark, E. Margaret Warton, Mary E. Reed, Stephanie Prausnitz, Jeffrey A. East, Mark F. Moeller, and Tracy A. Lieu. 2021. https://doi.org/10.1093/jamia/ocaa229 Physicians’ electronic inbox work patterns and factors associated with high inbox work duration . Journal of the American Medical Informatics Association, 28(5):923--930

work page doi:10.1093/jamia/ocaa229 2021

[2] [2]

Ayers, Adam Poliak, Mark Dredze, and et al

John W. Ayers, Adam Poliak, Mark Dredze, and et al. 2023. https://doi.org/10.1001/jamainternmed.2023.1838 Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum . JAMA Internal Medicine, 183(6):589--596. Published by the American Medical Association

arXiv 2023

[3] [3]

John W. et al. Ayers. 2023. Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine, 183(6):589--596

work page 2023

[4] [4]

Brands, S

M. Brands, S. Gouw, M. Beestrum, R. Cronin, K. Fijnvandraat, and S. Badawy. 2022. https://doi.org/10.2196/43086 Patient-centered digital health records and their effects on health outcomes: Systematic review . Journal of Medical Internet Research, 24(12):e43086

work page doi:10.2196/43086 2022

[5] [5]

Callahan

Alexander et al. Callahan. 2021. https://doi.org/10.1056/CAT.21.0224 Using aggregate patient data at the bedside via an on-demand consultation service . NEJM Catalyst Innovations in Care Delivery, 2

doi:10.1056/cat.21.0224 2021

[6] [6]

Carini, L

E. Carini, L. Villani, A. M. Pezzullo, A. Gentili, A. Barbara, W. Ricciardi, and S. Boccia. 2021. https://doi.org/10.2196/26189 The impact of digital patient portals on health outcomes, system efficiency, and patient attitudes: Updated systematic literature review . Journal of Medical Internet Research, 23(9):e26189

doi:10.2196/26189 2021

[7] [7]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024

[8] [8]

Yunxiang Li, Zihan Li, Kai Zhang, Ruilong Dan, Steve Jiang, and You Zhang. 2023. Chatdoctor: A medical chat model fine-tuned on a large language model meta-ai (llama) using medical domain knowledge. Cureus, 15(6)

2023