Pith. sign in

REVIEW 4 major objections 4 minor 41 references

ChiMed 2.0: Advancing Chinese Medical Dataset in Facilitating Large Language Modeling

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

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

Pith's one-line read Training Qwen-3 models through continued pre-training, supervised fine-tuning, and RLHF on a new 204.4M-character Chinese medical corpus yields accuracy gains on the medical subsets of the CMMLU and CEval benchmarks.

desk verdict Useful dataset, thin validation: the gains on CMMLU/CEval are plausible but a contamination audit is the missing piece before the claims hold. read the letter →

arxiv 2507.15275 v1 pith:CU2BUEJW submitted 2025-07-21 cs.CL

Yuanhe Tian , Junjie Liu , Zhizhou Kou , Yuxiang Li , Yan Song This is my paper
classification cs.CL
keywords Chinesemedicaldatasetlargelanguagemodelcontinuedpre-trainingsupervisedfine-tuningRLHFtraditionalmedicineCMMLUCEval
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 builds ChiMed 2.0, a 204.4-million-character Chinese medical corpus assembled from public medical websites and from LLM-generated content, spanning both traditional Chinese medicine classics and modern general clinical text. Its purpose is to show that one dataset can serve the entire modern LLM training regime: continued pre-training on 164.8K documents, supervised fine-tuning on 351.6K question-answer pairs, and RLHF on 41.7K preference tuples. The authors claim that running all three stages on Qwen-3 models of 1.7B and 14B parameters improves accuracy on the medical subsets of the CMMLU and CEval benchmarks, with larger relative gains at the smaller scale. If the claim holds, it matters because existing Chinese medical datasets are mostly QA pairs built for fine-tuning only, leaving pre-training and preference data as gaps this corpus fills for a high-demand but under-resourced language domain.

What carries the argument

The carrying artifact is the three-part corpus produced by a multi-stage pipeline. Raw text from five public Chinese medical sites, roughly 221M characters, is deduplicated, stripped of HTML and URLs, filtered by removing the top 1% of samples by Chinese GPT-2 perplexity, and cut to samples of at least 30 tokens. An LLM then does four advanced jobs: sensitive-content screening that deletes personal identifiers and toxic or misleading claims; automatic translation of classical Chinese medical passages into modern Chinese; generation of roughly 220K question-answer pairs from documents; and construction of preference data by generating two answers per question and having a second LLM pass rank one as accepted and the other as rejected. These steps yield 164.8K pre-training documents, 351.6K SFT pairs, and 41.7K RLHF tuples totaling 204.4M characters, with traditional medicine content (69.2M characters) balanced against modern clinical content (135.2M characters). Training itself uses LoRA on Qwen-3 base models for two pre-training epochs, two SFT epochs, and one RLHF epoch, with accuracy computed by the LM Evaluation Harness.

What would settle it

An overlap audit: compare the ChiMed 2.0 pre-training documents, SFT pairs, and preference tuples against the CMMLU and CEval medical test questions and answer options, looking for exact or near-duplicate n-grams and answer strings. If test content appears in the corpus, remove every overlapping sample, retrain the same Qwen-3 models with the same hyperparameters, and see whether the accuracy gains survive; if they shrink toward zero, memorization rather than medical ability explains the reported results.

Watch

Extended reading notes

Core claim

The central claim is that training on ChiMed 2.0 makes Chinese medical LLMs measurably more accurate. On Qwen-3 (1.7B), accuracy rises from 48.30 to 51.14 on CMMLU Genetics, from 60.07 to 61.17 on CMMLU College Medicine, and from 59.18 to 63.27 on CEval Physician; on Qwen-3 (14B), it rises from 87.55 to 89.01 and from 85.71 to 87.76 on the first and third of those benchmarks, while the College Medicine score stays at 76.70. The authors read the pattern of gains at both parameter scales as evidence that the corpus transfers medical knowledge rather than merely fitting a single model size, and they present the larger relative gains of the smaller model as a practical advantage for resource-constrained medical deployment. The dataset is organized so that each training stage consumes its own component, documents for pre-training, QA pairs for SFT, accepted and rejected answer pairs for RLHF, which is what lets a single corpus stand in for three separate resources.

Load-bearing premise

The gains count as real medical improvement only if the public website text used for training does not overlap with the public exam questions used for testing, and the paper does not check for that overlap.

Editorial extensions

If this is right

  • One dataset can plausibly cover the full training regime, pre-training, SFT, and RLHF, for a domain model, removing the need to stitch together separate corpora for each stage.
  • Smaller models gain more in relative accuracy, so teams with limited compute can still lift medical-domain performance meaningfully.
  • Automated translation makes classical Chinese medicine texts learnable by modern LLMs, folding a centuries-old knowledge base into the same training run as contemporary clinical text.
  • Because the pipeline combines public web scraping with LLM-driven cleaning and augmentation, the same construction recipe can be transplanted to other under-resourced medical languages.
  • The paper positions the corpus as supporting downstream tasks beyond benchmarks, including diagnostic assistance, clinical decision support, and medical question answering, on the strength of the benchmark gains.

Reading between the lines

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

  • My inference: the results do not rule out test-set contamination, because both the training text (public Chinese medical websites) and the evaluation sets (public CMMLU and CEval items) are drawn from the public web and no overlap check is reported; an n-gram overlap audit would settle this.
  • My reading: the claim of consistent gains across scales rests on five of six comparisons, since the Qwen-3 14B College Medicine score is unchanged at 76.70; the smaller-model results are the stronger evidence.
  • A testable extension the paper does not run: evaluate the trained models on open-ended medical consultation tasks or clinical assertion checking rather than multiple-choice exams, where memorized test items cannot inflate accuracy.
  • The QA pairs and preference rankings are generated by an unspecified LLM with no reported human audit of a sample; an expert review of a random subset would quantify how much generator bias the corpus passes on to the trained models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces ChiMed 2.0, a Chinese medical dataset with 204.4M characters organized into three parts: 164.8K documents for continued pretraining, 351.6K QA pairs for supervised fine-tuning (SFT), and 41.7K preference tuples for RLHF. The data are collected from public Chinese medical websites and processed with a multi-stage pipeline that includes cleaning, deduplication, ancient-to-modern translation, LLM-generated QA pairs, and LLM-ranked preference pairs. To validate the dataset, the authors continue pretraining, then SFT, then RLHF on Qwen-3 1.7B and 14B, and evaluate on the Genetics and College Medicine subsets of CMMLU and the physician subset of CEval. Table 6 reports gains in five of six cells, with one tie.

Significance. If the reported results hold, ChiMed 2.0 would be a useful resource because it is one of the few Chinese medical datasets that explicitly supports all three training stages, and it combines TCM classics with modern general medical content. The paper's strengths include detailed dataset statistics, disclosure of the processing prompts in Table 2, use of external benchmarks (CMMLU, CEval) rather than a self-defined metric, and experiments at two model scales. The main weakness is that the reported gains rest on a small number of benchmark cells with no contamination audit and no statistical robustness checks, so the central contribution bullet is conditional on those checks coming out clean.

major comments (4)
  1. [§3.1, §4.1, Table 6] The central claim that ChiMed 2.0 improves CMMLU and CEval accuracy is vulnerable to benchmark contamination: the corpus is scraped from public Chinese medical websites, and the evaluation uses public Chinese benchmark subsets, yet no overlap check between the corpus and the benchmark items is reported anywhere in Sections 3.2–4.2. Because CMMLU and CEval questions draw on Chinese medical examination content that is often reproduced on public health and QA sites, exact or near-duplicate items in the pretraining documents or generated QA pairs could allow answer memorization. Please add a decontamination audit (e.g., n-gram overlap, embedding similarity, and manual inspection of near-duplicates) against the exact benchmark items used, and report the results separately for the pretraining, SFT, and preference data. If overlap is found, decontaminate the corpus and rerun Table 6.
  2. [§4.2, Table 6] The 'consistent performance gains' claim rests on six cells, and one of them (CMMLU College Medicine with Qwen-3 14B) shows no gain (76.70 vs. 76.70). The paper reports no error bars, multiple seeds, or significance tests. Please report means and standard deviations over at least three runs/seeds for both the original and trained models, or otherwise show that the observed gains exceed run-to-run variation; without this, the 1.10–4.09 percentage point gains cannot be distinguished from noise.
  3. [§3.2, Table 2] The RLHF preference data is generated and ranked entirely by LLMs: one LLM writes two answers to the same question and another LLM selects the better one, with no human validation or agreement statistics. The paper therefore assumes, without evidence, that model judgments substitute for human preferences in RLHF. Please provide a human-annotation study on a random sample of preference tuples (e.g., agreement rate between the LLM judge and human raters), state which LLM was used for ranking, and, if possible, include an ablation that trains with and without the preference data.
  4. [§3.2, §4.1] No ablation separates the contributions of continued pretraining, SFT, and RLHF. Since Table 6 reports only the end-to-end pipeline, it is unclear whether the gains come from the ChiMed 2.0 corpus specifically, from one of the three stages, or from any additional training on domain text. Please report intermediate results after each stage (pretraining only, +SFT, +RLHF) for each model and benchmark subset.
minor comments (4)
  1. [§4.2] The text refers to 'CEval Physics subsets', but Section 4.1 defines the subset as the physician (Phy) subset of CEval; please fix this typo to avoid confusion.
  2. [§3.2] The paper does not state which LLM was used for sensitive-content cleaning, QA generation, translation, and preference ranking. Identifying the model (and version) is essential for reproducibility and for assessing the risk of self-generated data contamination.
  3. [§4.1] LoRA is used for efficient training, but the LoRA rank, alpha, and dropout are not reported. Please add these hyperparameters to Table 5 or the surrounding text.
  4. [§4.1] The number of test questions in each CMMLU/CEval subset is not reported. Since the subsets may be small, please report the sample sizes so readers can interpret the accuracy differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is validated against external benchmarks, not against the dataset's own construction.

full rationale

The paper's load-bearing empirical claim is that continued pretraining, SFT, and RLHF on ChiMed 2.0 improve Qwen-3 models on CMMLU Genetics and College Medicine subsets and the CEval physician subset (Table 6). Those benchmarks are external, public multiple-choice suites that are not constructed from, defined by, or fitted to the ChiMed 2.0 corpus. The dataset construction steps (web crawling, cleaning, perplexity filtering, LLM-based translation, QA generation, and preference generation) do not define the evaluation accuracy, and no benchmark score is used as a training target or as a parameter in the data pipeline. The claimed gain is therefore not equivalent to an input of the pipeline by construction. The only associated concern is possible benchmark contamination from public medical websites, but that is an empirical validity risk requiring an overlap audit, not a circularity detectable in the derivation itself. The paper's self-citations (ChiMed as a predecessor and ChiMed-GPT as related work) are provenance references and do not carry the load-bearing argument: the improvement is attributed to training on ChiMed 2.0 and measured on independent benchmarks. Accordingly, no circular step can be exhibited with the required specificity, and the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The dataset is a curation of existing texts plus LLM-generated QA and preference data, so the main burden falls on the quality of LLM outputs and the independence of the benchmark evaluations.

free parameters (2)
  • Perplexity filtering threshold = top 1% PPL removed
    Samples with the highest Chinese GPT-2 perplexity are discarded; the top-1% cutoff is hand-chosen and not validated (Section 3.2).
  • Minimum text length = 30 tokens
    Samples shorter than 30 tokens are removed; this hand-chosen cutoff affects which medical texts survive preprocessing (Section 3.2).
assumptions (5)
  • domain assumption LLM-generated QA pairs and preference rankings are accurate enough to serve as training signals
    Section 3.2 describes QA and preference data produced by LLMs without human verification; the SFT and preference-training experiments rely on this assumption.
  • domain assumption Chinese GPT-2 perplexity is a valid quality filter for Chinese medical text
    A general-domain Chinese GPT-2 is used to remove high-PPL samples (Section 3.2); there is no validation that this filtering preserves medically correct content.
  • domain assumption CMMLU and CEval medical multiple-choice accuracy measures meaningful medical capability
    Evaluation uses only top-prediction accuracy on three benchmark subsets (Section 4.1); no calibration, expert review, or robustness analysis is provided.
  • domain assumption ChiMed 2.0 does not substantially overlap with CMMLU/CEval test questions
    Data is scraped from public Chinese medical websites (Section 3.1) and evaluated on public benchmarks (Section 4.1), but no contamination analysis is reported.
  • ad hoc to paper LLM-ranked preferences substitute for human feedback in RLHF
    Preference tuples are built by having an LLM rank two LLM-generated answers (Table 2), with no human annotators, yet the paper labels this stage RLHF (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChiMed 2.0: Advancing Chinese Medical Dataset in Facilitating Large Language Modeling." pith.science (2026). https://pith.science/paper/CU2BUEJW

@misc{pith2026250715275,
  author       = {Pith},
  title        = {Pith review of: ChiMed 2.0: Advancing Chinese Medical Dataset in Facilitating Large Language Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CU2BUEJW}},
  note         = {Machine review of arXiv:2507.15275}
}
read the original abstract

Building high-quality data resources is crucial for advancing artificial intelligence research and applications in specific domains, particularly in the Chinese medical domain. Existing Chinese medical datasets are limited in size and narrow in domain coverage, falling short of the diverse corpora required for effective pre-training. Moreover, most datasets are designed solely for LLM fine-tuning and do not support pre-training and reinforcement learning from human feedback (RLHF). In this paper, we propose a Chinese medical dataset named ChiMed 2.0, which extends our previous work ChiMed, and covers data collected from Chinese medical online platforms and generated by LLMs. ChiMed 2.0 contains 204.4M Chinese characters covering both traditional Chinese medicine classics and modern general medical data, where there are 164.8K documents for pre-training, 351.6K question-answering pairs for supervised fine-tuning (SFT), and 41.7K preference data tuples for RLHF. To validate the effectiveness of our approach for training a Chinese medical LLM, we conduct further pre-training, SFT, and RLHF experiments on representative general domain LLMs and evaluate their performance on medical benchmark datasets. The results show performance gains across different model scales, validating the dataset's effectiveness and applicability.

Figures

Figures reproduced from arXiv: 2507.15275 by the authors.

Figure 1
Figure 1. The distribution of the department of the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overall data processing process. analysis, visual processing techniques, and AI￾driven compatibility alerts. • China Health Network4 : As a nationally renowned comprehensive health portal in China, this platform offers practical services in￾cluding online consultations with top-tier hospi￾tal physicians, health self-assessment tools, and medical care navigation guidance. Its website content is principally scrape… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 25 canonical work pages

  1. [1]

    Asma Ben Abacha, Yassine Mrabet, Mark Sharp, Travis R Goodwin, Sonya E Shooshan, and Dina Demner-Fushman. 2019. Bridging the gap between consumers’ medication questions and trusted answers. In MEDINFO 2019: Health and Wellbeing e-Networks for All, pages 25--29. IOS Press

    work page 2019

  2. [2]

    Zahra Abbasiantaeb, Yifei Yuan, Evangelos Kanoulas, and Mohammad Aliannejadi. 2024. Let the llms talk: Simulating human-to-human conversational qa via zero-shot llm-to-llm interactions. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 8--17

    work page 2024

  3. [3]

    Hanjie Chen, Zhouxiang Fang, Yash Singla, and Mark Dredze. 2025. Benchmarking large language models on answering and explaining challenging medical questions. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3563--3599

    work page 2025

  4. [4]

    Xiongtao Cui and Jungang Han. 2020. Chinese medical question answer matching based on interactive sentence representation learning. arXiv preprint arXiv:2011.13573

    work page Pith review arXiv 2020

  5. [5]

    Siyuan Dai, Kai Ye, Guodong Liu, Haoteng Tang, and Liang Zhan. 2025. Zeus: Zero-shot llm instruction for union segmentation in multimodal medical imaging. arXiv preprint arXiv:2504.07336

    work page Pith review arXiv 2025

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT : Pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171--4186, Minneapolis, Minnesota

    work page 2019

  7. [7]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2024. The l...

    work page 2024

  8. [8]

    Tongfeng Guan, Hongying Zan, Xiabing Zhou, Hongfei Xu, and Kunli Zhang. 2020. Cmeie: Construction and evaluation of chinese medical information extraction dataset. In Natural Language Processing and Chinese Computing: 9th CCF International Conference, NLPCC 2020, Zhengzhou, China, October 14--18, 2020, Proceedings, Part I 9, pages 270--282. Springer

    work page 2020

Show all 41 references
  1. [9]

    Jialong Han, Yan Song, Wayne Xin Zhao, Shuming Shi, and Haisong Zhang. 2018. hyperdoc2vec: Distributed representations of hypertext documents. arXiv preprint arXiv:1805.03793

    2018 arXiv

  2. [10]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  3. [11]

    Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. 2023. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems...

    2023

  4. [12]

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

    2021

  5. [13]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W Cohen, and Xinghua Lu. 2019. Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146

    2019 arXiv

  6. [14]

    Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1--9

    2016

  7. [15]

    Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2024 a . CMMLU : Measuring massive multitask language understanding in C hinese. In Findings of the Association for Computational Linguistics: ACL 2024, pages 11260--11285, Ba...

    2024

  8. [16]

    Shuyue Stella Li, Vidhisha Balachandran, Shangbin Feng, Jonathan Ilgen, Emma Pierson, Pang Wei Koh, and Yulia Tsvetkov. 2024 b . Mediq: Question-asking llms for adaptive and reliable medical reasoning. arXiv e-prints, pages arXiv--2406

    2024

  9. [17]

    Junling Liu, Peilin Zhou, Yining Hua, Dading Chong, Zhongyu Tian, Andrew Liu, Helin Wang, Chenyu You, Zhenhua Guo, Lei Zhu, et al. 2023 a . Benchmarking large language models on cmexam-a comprehensive chinese medical exam dataset. Advances in Neural Information Processing Syst...

    2023

  10. [18]

    Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, and Daochen Zha. 2023 b . Fingpt: Democratizing internet-scale data for financial large language models. arXiv preprint arXiv:2307.10485

    2023 arXiv

  11. [19]

    Masoud Monajatipoor, Jiaxin Yang, Joel Stremmel, Melika Emami, Fazlolah Mohaghegh, Mozhdeh Rouhsedaghat, and Kai-Wei Chang. 2024. Llms in biomedicine: A study on clinical named entity recognition. arXiv preprint arXiv:2404.07376

    2024 arXiv

  12. [20]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training L anguage M odels to F ollow I nstructions with H uman F eedback. Advances in Neural Information Processing Syst...

    2022

  13. [21]

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. 2022. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Conference on health, inference, and learning, pages 248--260. PMLR

    2022

  14. [22]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

    2023

  15. [23]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  16. [24]

    Yan Song and Shuming Shi. 2018. Complementary L earning of W ord E mbeddings. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18 , pages 4368--4374

    2018

  17. [25]

    Yan Song, Yuanhe Tian, Nan Wang, and Fei Xia. 2020. Summarizing M edical C onversations via I dentifying I mportant U tterances. In Proceedings of the 28th International Conference on Computational Linguistics, pages 717--729

    2020

  18. [26]

    Yan Song, Tong Zhang, Yonggang Wang, and Kai-Fu Lee. 2021. ZEN 2.0: C ontinue T raining and A daption for N -gram E nhanced T ext E ncoders. arXiv preprint arXiv:2105.01279

    2021 arXiv

  19. [27]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford Alpaca: An Instruction-following LLaMA model . GitHub repository

    2023

  20. [28]

    Yuanhe Tian, Ruyi Gan, Yan Song, Jiaxing Zhang, and Yongdong Zhang. 2024 a . C hi M ed- GPT : A C hinese medical large language model with full training regime and better alignment to human preferences. In Proceedings of the 62nd Annual Meeting of the Association for Computati...

    2024

  21. [29]

    Yuanhe Tian, Weicheng Ma, Fei Xia, and Yan Song. 2019. Chimed: A chinese medical corpus for question answering. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 250--260

    2019

  22. [30]

    Yuanhe Tian, Chen Su, Junwen Duan, and Yan Song. 2025. Computed tomography visual question answering with cross-modal feature graphing. arXiv preprint arXiv:2507.04333

    2025 arXiv

  23. [31]

    Yuanhe Tian, Fei Xia, and Yan Song. 2024 b . Dialogue summarization with mixture of experts based on large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7143--7155

    2024

  24. [32]

    Yuanhe Tian, Fei Xia, and Yan Song. 2024 c . Diffusion networks with task-specific noise control for radiology report generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1771--1780

    2024

  25. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and Efficient Foundation Language Models . arXiv preprint arXiv:2302.13971

    2023 arXiv

  26. [34]

    Guanchu Wang, Junhao Ran, Ruixiang Tang, Chia-Yuan Chang, Yu-Neng Chuang, Zirui Liu, Vladimir Braverman, Zhandong Liu, and Xia Hu. 2024. Assessing and enhancing large language models in rare disease question-answering. arXiv preprint arXiv:2408.08422

    2024

  27. [35]

    Xidong Wang, Guiming Hardy Chen, Dingjie Song, Zhiyi Zhang, Zhihong Chen, Qingying Xiao, Feng Jiang, Jianquan Li, Xiang Wan, Benyou Wang, et al. 2023. Cmb: A comprehensive medical benchmark in chinese. arXiv preprint arXiv:2308.08833

    2023 arXiv

  28. [36]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  29. [37]

    Songhua Yang, Hanjie Zhao, Senbin Zhu, Guangyu Zhou, Hongfei Xu, Yuxiang Jia, and Hongying Zan. 2024. Zhongjing: Enhancing the chinese medical capabilities of large language model through expert feedback and real-world multi-turn dialogue. In Proceedings of the AAAI conference...

    2024

  30. [38]

    Dong Yuan, Eti Rastogi, Gautam Naik, Sree Prasanna Rajagopal, Sagar Goyal, Fen Zhao, Bharath Chintagunta, and Jeff Ward. 2024. A continued pretrained llm approach for automatic medical note generation. arXiv preprint arXiv:2403.09057

    2024 arXiv

  31. [39]

    Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, Hongchao Fang, Penghui Zhu, Shu Chen, and Pengtao Xie. 2020. M ed D ialog: Large-scale medical dialogue datasets. In Proceedings of the 2020 Conferenc...

    2020

  32. [40]

    Zirui Zhao, Wee Sun Lee, and David Hsu. 2023. Large Language Models as Commonsense Knowledge for Large-Scale Task Planning . arXiv preprint arXiv:2305.14078

    2023 arXiv

  33. [41]

    Zhi Zhou, Jiang-Xin Shi, Peng-Xiao Song, Xiao-Wen Yang, Yi-Xuan Jin, Lan-Zhe Guo, and Yu-Feng Li. 2024. Lawgpt: A chinese legal knowledge-enhanced large language model. arXiv preprint arXiv:2406.04614

    2024 arXiv

Pith tools

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