pith. sign in

arxiv: 2605.24635 · v1 · pith:46BFV75Xnew · submitted 2026-05-23 · 💻 cs.CL

HiMed: Incentivizing Hindi Reasoning in Medical LLMs

Dingfeng Jiang , Han Yan , Chenze Ma , Amit Kumar Jaiswal , Ang Li , Yunxiang Jiang , Xinlei Xiong , Juhao Liang
show 7 more authors
Hongru Xiao Xiang Li Fan Bu Jiale Han Ruchir Gupta Prayag Tiwari Benyou Wang
This is my paper

Pith reviewed 2026-06-30 13:27 UTC · model grok-4.3

classification 💻 cs.CL
keywords Hindi medical reasoningmedical LLMscross-lingual transferscaffolding rewardHiMed corpusIndian systems of medicinelanguage model fine-tuninghealthcare disparities
0
0 comments X

The pith

A new Hindi medical corpus and decaying scaffolding reward improve LLM reasoning in Hindi and shrink the English performance gap.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to establish that medical LLMs suffer sharp drops in Hindi, especially on Indian systems of medicine, and that this can be addressed by building a dedicated Hindi reasoning corpus and benchmark called HiMed that covers both Western and Indian medicine. The authors train an 8B model, HiMed-8B, with a decaying scaffolding reward that gradually reduces external guidance to force genuine reasoning steps. Experiments show gains in Hindi accuracy and a narrower English-Hindi gap, with ablations confirming that each training stage and reward element contributes. A sympathetic reader would care because closing the language gap could extend reliable medical AI support to Hindi-speaking populations and reduce healthcare disparities in India.

Core claim

The paper claims that constructing the HiMed corpus and benchmark, then training with a decaying scaffolding reward, produces a model (HiMed-8B) that raises Hindi medical reasoning performance across Western and Indian medicine topics while measurably reducing the English-Hindi accuracy difference.

What carries the argument

The HiMed corpus and benchmark suite, combined with the decaying scaffolding reward that starts with heavy guidance and gradually withdraws it to promote independent reasoning.

If this is right

  • Hindi-speaking patients and clinicians gain access to more reliable medical LLM outputs on both allopathic and traditional Indian medicine topics.
  • Similar corpus-plus-reward pipelines could be applied to other low-resource languages for medical or technical domains.
  • Ablation results indicate that removing any single training stage or reward component measurably weakens the Hindi gains.
  • The approach narrows cross-lingual gaps without requiring full retraining from scratch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the decaying reward generalizes, the same schedule might improve reasoning in non-medical low-resource language settings.
  • The HiMed construction method supplies a reusable template for building bilingual medical benchmarks that include indigenous medical systems.
  • Persistent gaps after this intervention would point to deeper architectural limits on cross-lingual transfer rather than data scarcity alone.

Load-bearing premise

The HiMed corpus and benchmark give a faithful, unbiased measure of real-world Hindi medical reasoning ability, and the decaying reward creates genuine reasoning gains instead of benchmark-specific overfitting.

What would settle it

An external test set of Hindi medical questions drawn from practicing Indian physicians and not used in HiMed training shows no accuracy gain or no reduction in the English-Hindi gap.

Figures

Figures reproduced from arXiv: 2605.24635 by Amit Kumar Jaiswal, Ang Li, Benyou Wang, Chenze Ma, Dingfeng Jiang, Fan Bu, Han Yan, Hongru Xiao, Jiale Han, Juhao Liang, Prayag Tiwari, Ruchir Gupta, Xiang Li, Xinlei Xiong, Yunxiang Jiang.

Figure 1
Figure 1. Figure 1: Overview of HiMed and the training framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of rewards. 4.5 Reliability of the Accuracy Reward Model We fine-tune LLaMA-3.2-3B-Instruct on a bilin￾gual medical reasoning dataset as an accuracy veri￾fier. The probability of the True class from the soft￾max over binary logits is used as the confidence score. On the test set, the verifier achieves 0.969 precision and 0.936 recall. To assess reliability, we use GPT-5 to generate 300 ch… view at source ↗
Figure 3
Figure 3. Figure 3: OCR manual inspection UI. D.2 Instruction Pool The instruction pool defines a standardized set of medical question templates used to guide instruc￾tion generation from OCR-extracted text. It serves as an intermediate layer between raw medical pas￾sages and downstream data generation, bridging the unstructured source content with the structured instruction formats. Method. The pool is constructed from high￾… view at source ↗
read the original abstract

Medical large language models hold promise for reducing healthcare disparities, yet Hindi remains severely underrepresented. While medical LLMs excel in high-resource languages, their performance degrades sharply in Hindi, particularly on Indian systems of medicine. We argue that robust cross-lingual medical transfer requires Hindi reasoning. To this end, we introduce HiMed, a Hindi reasoning medical corpus and benchmark suite covering both Western and Indian medicine. We further propose HiMed-8B, a Hindi-form medical reasoning LLM, through the design of decaying scaffolding reward. Extensive experiments demonstrate improvement in Hindi medical reasoning performance and reduction in the English--Hindi accuracy gap. Ablation studies validate the contribution of each training stage and reward component. All data and code are available on GitHub: https://github.com/FreedomIntelligence/HiMed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces HiMed, a new Hindi medical reasoning corpus and benchmark covering both Western and Indian medicine, along with HiMed-8B, an 8B-parameter model trained via a decaying scaffolding reward. It claims that this training produces improved Hindi medical reasoning performance and a reduced English–Hindi accuracy gap, with ablation studies validating each component; all data and code are released publicly.

Significance. If the benchmark is shown to be uncontaminated and the gains prove transferable, the work would address a clear gap in multilingual medical LLMs by targeting an underrepresented language and including Indian systems of medicine. Public release of the corpus, benchmark, and code is a concrete strength that supports reproducibility and follow-on research.

major comments (3)
  1. [Abstract] Abstract: the claim of 'extensive experiments demonstrate improvement in Hindi medical reasoning performance and reduction in the English--Hindi accuracy gap' is unsupported by any quantitative results, baselines, dataset sizes, error bars, or evaluation protocols. Without these numbers the central empirical claim cannot be assessed.
  2. [Benchmark construction] Benchmark construction section: no information is supplied on question sourcing, difficulty filtering, overlap testing with training data, or checks for translation artifacts. This detail is load-bearing for the claim that observed gains reflect genuine Hindi reasoning rather than benchmark-specific artifacts or contamination.
  3. [Ablation studies / training procedure] Ablation and training sections: the abstract states that ablations validate each training stage and reward component, yet no controls are described (e.g., held-out non-HiMed Hindi medical tasks or plain SFT without the decaying reward schedule). This omission prevents evaluation of whether the scaffolding reward produces transferable reasoning gains.
minor comments (1)
  1. [Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., accuracy delta or gap reduction) so readers can immediately gauge effect size.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to improve clarity and completeness where needed.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'extensive experiments demonstrate improvement in Hindi medical reasoning performance and reduction in the English--Hindi accuracy gap' is unsupported by any quantitative results, baselines, dataset sizes, error bars, or evaluation protocols. Without these numbers the central empirical claim cannot be assessed.

    Authors: We agree that the abstract should include key quantitative results to support the central claims. In the revised version, we will add specific performance numbers (e.g., accuracy improvements on Hindi tasks), baseline comparisons, dataset sizes, error bars where applicable, and a brief description of the evaluation protocol. revision: yes

  2. Referee: [Benchmark construction] Benchmark construction section: no information is supplied on question sourcing, difficulty filtering, overlap testing with training data, or checks for translation artifacts. This detail is load-bearing for the claim that observed gains reflect genuine Hindi reasoning rather than benchmark-specific artifacts or contamination.

    Authors: The referee correctly notes that these methodological details are insufficiently described. We will revise the benchmark construction section to explicitly detail question sourcing (from Hindi medical textbooks, exam papers, and Indian health resources), difficulty filtering (via expert annotation), overlap testing (n-gram and embedding-based checks against training data), and translation artifact checks (back-translation and native-speaker validation). revision: yes

  3. Referee: [Ablation studies / training procedure] Ablation and training sections: the abstract states that ablations validate each training stage and reward component, yet no controls are described (e.g., held-out non-HiMed Hindi medical tasks or plain SFT without the decaying reward schedule). This omission prevents evaluation of whether the scaffolding reward produces transferable reasoning gains.

    Authors: We agree that additional controls are needed to strengthen the ablation analysis. We will expand the ablation and training sections to include results on held-out non-HiMed Hindi medical tasks and direct comparisons to plain SFT without the decaying scaffolding reward, to better demonstrate the contribution of each component. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims rest on experiments, not derivations reducing to inputs

full rationale

The paper introduces a new corpus/benchmark (HiMed) and trains HiMed-8B using a decaying scaffolding reward, then reports experimental improvements in Hindi medical reasoning and gap reduction. No equations, first-principles derivations, fitted parameters renamed as predictions, or self-citation load-bearing steps appear in the provided text or abstract. Ablations are described as validating components, but these are standard experimental controls rather than reductions to the paper's own inputs by construction. The central claims are therefore self-contained against external benchmarks and receive the default non-finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical derivations, free parameters, axioms, or invented entities are mentioned in the abstract; the work is empirical, centered on new data creation and a training procedure.

pith-pipeline@v0.9.1-grok · 5708 in / 1116 out tokens · 47382 ms · 2026-06-30T13:27:18.262334+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    arXiv preprint arXiv:2508.14828

    Long chain-of-thought reasoning across lan- guages. arXiv preprint arXiv:2508.14828. Linzheng Chai, Jian Y ang, Tao Sun, Hongcheng Guo, Jiaheng Liu, Bing Wang, Xinnian Liang, Jiaqi Bai, Tongliang Li, Qiyao Peng, and Zhoujun Li. 2025. XCOT: cross-lingual instruction tuning for cross- lingual chain-of-thought reasoning . In Thirty-Ninth AAAI Conference on A...

    work page arXiv 2025

  2. [2]

    In Find- ings of the Association for Computational Linguis- tics: ACL 2025 , pages 14552–14573, Vienna, Aus- tria

    Towards medical complex reasoning with LLMs through medical verifiable problems . In Find- ings of the Association for Computational Linguis- tics: ACL 2025 , pages 14552–14573, Vienna, Aus- tria. Association for Computational Linguistics. Sribala Vidyadhari Chinta, Zichong Wang, Xingyu Zhang, Thang Doan Viet, Ayesha Kashif, Monique Antoinette Smith, and ...

    2025

  3. [3]

    arXiv preprint arXiv:2407.19655

    Ai-driven healthcare: A survey on ensur- ing fairness and mitigating bias . arXiv preprint arXiv:2407.19655. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry T worek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman

    work page arXiv

  4. [4]

    Training Verifiers to Solve Math Word Problems

    Training verifiers to solve math word prob- lems. CoRR, abs/2110.14168. Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Mail- lard, Anna Y . Sun, Skyler Wang, Guillaume Wen- zek, Al Y oungblood, Bapi Akula, Loïc Barrault, Gabriel Mejia Gonzalez, Prangthip Hansa...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    CoRR, abs/2510.25409

    Bhashabench V1: A comprehensive bench- mark for the quadrant of indic domains . CoRR, abs/2510.25409. Government of NCT of Delhi Directorate of AYUSH. n.d. Ashtāṇga ayurveda overview. https://ayush. delhi.gov.in/ayush/ayurveda. Accessed 2025-09- 20. John W Ely, Jerome A Osheroff, Mark H Ebell, George R Bergus, Barcey T Levy, M Lee Chamb- liss, and Eric R ...

    work page arXiv 2025

  6. [6]

    Danni Liu and Jan Niehues

    Let’s verify step by step . Danni Liu and Jan Niehues. 2025. Middle-layer repre- sentation alignment for cross-lingual transfer in fine- tuned LLMs . In Proceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 15979– 15996, Vienna, Austria. Association for Computa- tional Linguistics. Y ang ...

    2025

  7. [7]

    InInternational Conference on Learning Representations, volume 2024, pages 39578–39601

    Biogpt: Generative pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics. Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Y ang Y ou, and Hongxia Y ang. 2023. Let’s reward step by step: Step-level reward model as the navigators for reasoning . CoRR, abs/2310.10080. Sharmistha Mallick. 2016. Challeng...

    work page arXiv 2023

  8. [8]

    World Health Organization

    Cure-med: Curriculum-informed reinforce- ment learning for multilingual medical reasoning . World Health Organization. 2022. Who benchmarks for training in yoga . In development; see WHO Tradi- tional Medicine programme updates. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina...

    work page arXiv 2022

  9. [9]

    arXiv preprint arXiv:2411.02538

    Milu: A multi-task indic language understand- ing benchmark. arXiv preprint arXiv:2411.02538. Veniamin Veselovsky, Berke Argin, Benedikt Stroebl, Chris Wendler, Robert West, James Evans, Thomas L. Griffiths, and Arvind Narayanan

    work page arXiv

  10. [10]

    Griffiths, and Arvind Narayanan

    Localized cultural knowledge is conserved and controllable in large language models . CoRR, abs/2504.10191. L. S. Vygotsky. 1978. Mind in society: The develop- ment of higher psychological processes. Harvard University Press. 15 Guoxin Wang, Minyu Gao, Shuai Y ang, Y a Zhang, Lizhi He, Liang Huang, Hanlin Xiao, Y exuan Zhang, Wanyue Li, Lu Chen, Jintao Fe...

    work page arXiv 1978

  11. [11]

    arXiv preprint arXiv:2411.14461

    Towards next-generation medical agent: How o1 is reshaping decision-making in medical scenar- ios. arXiv preprint arXiv:2411.14461. Wen Y ang, Junhong Wu, Chen Wang, Chengqing Zong, and Jiajun Zhang. 2025. Implicit cross-lingual re- warding for efficient multilingual preference align- ment. In Findings of the Association for Computa- tional Linguistics: AC...

    work page arXiv 2025

  12. [12]

    Tree of thoughts: Deliberate problem solv- ing with large language models. In Advances in Neu- ral Information Processing Systems 36: Annual Con- ference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, De- cember 10 - 16, 2023 . Kaiyan Zhang, Sihang Zeng, Ermo Hua, Ning Ding, Zhang-Ren Chen, Zhiyuan Ma, Haoxin Li, Ganqu ...

    work page arXiv 2023

  13. [13]

    In The Thirteenth Inter- national Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025

    Judgelm: Fine-tuned large language mod- els are scalable judges . In The Thirteenth Inter- national Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenRe- view.net. Yuxin Zuo, Shang Qu, Yifei Li, Zhang-Ren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. 2025. Medxpertqa: Benchmark- ing expert-level m...

    2025

  14. [14]

    include Hindi but lack medical specializa- tion or rely on translated data. AyurGenixAI ( kag- glekirti123, 2025) covers Indian medicine yet remains largely English, while the Multilingual 17 Healthcare Text Dataset ( Bagga, 2025) provides diverse content without QA or MCQ structure. Furthermore, only MedMCQA, ReasonMed, and MedReason include CoT annotati...

    2025

  15. [15]

    Overall, existing benchmarks remain English-centric and do not sup- port comprehensive evaluation across both West- ern and Indian medicine

    consists primarily of OCR outputs from ex- aminations, but merely focuses on Ayurveda, leav- ing a huge gap on other streams. Overall, existing benchmarks remain English-centric and do not sup- port comprehensive evaluation across both West- ern and Indian medicine. Cultural alignment. Prior work addresses cul- tural grounding via region-specific data cur...

    2024

  16. [16]

    Despite these rapid ad- vances, medical reward modeling remains largely accuracy-centric, with limited consideration of lan- guage nativeness or cultural fidelity

    addresses reward imbalance by optimizing worst-case performance. Despite these rapid ad- vances, medical reward modeling remains largely accuracy-centric, with limited consideration of lan- guage nativeness or cultural fidelity. Medical Reasoning LLMs. Recent work has advanced medical reasoning in LLMs through domain-adapted models such as Med-PaLM2, BioG...

    2022

  17. [17]

    The ambiguity concerns a key subject/objec- t/symptom/treatment/causal relation and blocks correct medical interpretation

  18. [18]

    The ambiguous element cannot be resolved from the text itself, even with generous natural-language inference

  19. [19]

    The ambiguity spans the entire text (the text never provides enough information to clarify the refer- ent)

  20. [20]

    It is not a normal omission/abbreviation/vague- ness/stylistic shortening/common medical phras- ing

  21. [21]

    If the text is understandable overall (even with minor un- clear parts), then has_ambiguity must be False

    It is not a minor unclear phrase that does not affect the main meaning. If the text is understandable overall (even with minor un- clear parts), then has_ambiguity must be False. Rules: • Output only the 4 lines above. • No explanations. • Use Python-style booleans: True/False. Text: {text} Prompt: LLM Calibration Y ou are an OCR post-correction assistant...

  22. [22]

    • Minor typos, spacing issues, missing matras, imperfect Hindi, or small artifacts are not major OCR problems

    NO_PROBLEM: • The text is readable, understandable, and mostly coherent. • Minor typos, spacing issues, missing matras, imperfect Hindi, or small artifacts are not major OCR problems. • If a normal Hindi reader can infer the mean- ing easily, choose NO_PROBLEM

  23. [23]

    POSSIBLE_ISSUE: • One or two suspicious fragments may be OCR errors, but overall meaning is still un- derstandable

  24. [24]

    Output format: • Line 1: exactly one of NO_PROBLEM / POSSI- BLE_ISSUE / DEFINITE_ISSUE • Line 2: a short English explanation Do not output Markdown, JSON, or extra text

    DEFINITE_ISSUE: • The text is seriously corrupted (broken words, incomplete fragments, nonsense se- quences) and the meaning cannot be recov- ered. Output format: • Line 1: exactly one of NO_PROBLEM / POSSI- BLE_ISSUE / DEFINITE_ISSUE • Line 2: a short English explanation Do not output Markdown, JSON, or extra text. Text to evaluate: {text} Prompt: Formin...

  25. [25]

    • Identify and distill the user’s core medical consultation intent

    Information Analysis and Intent Extraction • Carefully read all information points pro- vided by the user, such as symptoms, history, test indicators, doctor‘s advice, demograph- ics, and location. • Identify and distill the user’s core medical consultation intent. 27

  26. [26]

    In the con- versation, the user provided [Informa- tion 1], [Information 2], ..., ultimately asking [Core Question]

    Standardized Output Generation • Based on question complexity, generate Instruction in one of the following formats: – A. Conversation Summary Format (multiple user utterances): Use a narrative summary: “In the con- versation, the user provided [Informa- tion 1], [Information 2], ..., ultimately asking [Core Question].” – B. QA Standard Format (one user q...

  27. [27]

    Intent Classification • Assign one primary label: Diagnosis, Treatment, Etiology, Prognosis, or Medical Knowledge

  28. [28]

    Semantic Deduplication • Identify instructions with identical core in- tent and retain only the clearest, most stan- dard one for each group

  29. [29]

    Final Output: Return the Markdown table only

    Output Format • Return a Markdown table with columns: Instruction, Category, Form. Final Output: Return the Markdown table only. Prompt:Template-Guided Instruction Gen- eration Role Y ou are an expert question writer for traditional medicine. Y our task is to generate high-quality instruction-style data (QA, MCQ, or Dialogue) grounded in traditional medic...

  30. [30]

    Select the given template and follow its structure strictly

  31. [31]

    Generate three difficulty levels: EASY, MEDIUM, and HARD

  32. [32]

    Ensure all outputs are self-contained, grounded in the source text, and consistent with traditional medicine

  33. [33]

    Malformed generations (e.g., missing MCQ options, schema violations, or empty rationales when required) will be automatically rejected

    Produce structured outputs that satisfy the required schema. Malformed generations (e.g., missing MCQ options, schema violations, or empty rationales when required) will be automatically rejected. Few-Shot Style References The following few-shot blocks are provided as style references only (do NOT copy sentences verbatim): • MCQ_FEW_SHOT {...} • QA_FEW_SH...

  34. [34]

    grounded_in_context : • Score 1.00 if all information in question/answer/cot is directly derivable from the source text • Score lower if there is any hallucinated information or external knowledge not present in the text • Score 0.00 if the content is completely unrelated to the source text

  35. [35]

    medical_correctness : • Score 1.00 if the medical information is correct according to the source text • Score lower if there are minor inaccuracies or misinterpretations • Score 0.00 if the medical information is incorrect or contradicts the source text

  36. [36]

    reasoning_clarity: • Score 1.00 if the reasoning steps are logical, clear, and well-structured • Score lower if reasoning is somewhat unclear or has minor logical gaps • Score 0.00 if reasoning is illogical, confusing, or missing

  37. [37]

    Return only the JSON object

    language_quality: • Score 1.00 if the Hindi language is fluent, natural, and grammatically correct • Score lower if there are minor grammatical errors or awkward phrasing • Score 0.00 if the language is severely broken or incomprehensible Return ONL Y a JSON object with 4 float scores in [0.00, 1.00]: { ”grounded_in_context”: 0.00, ”medical_correctness”: ...

  38. [38]

    Y ou have sudden blurred vision in your left eye

    Imagine you wake up feeling unwell. Y ou have sudden blurred vision in your left eye. Y ou decide to message a doctor or a reliable AI health assistant for advice. How would you describe your symptoms to a doc- tor or AI to get help with diagnosis? कल्‍पना कीि जए िक आप बीमा र महसूस करते हु ए उठते हैं । आपकी बाईं आंख म ें अचा नक ध ुंधला िदखाई देने लगा है। ...

  39. [39]

    मेर े पि र- णाम असामान्‍य क्‍यों हैं और इन बदलावों का का रण क्‍या हो सकता है?

    After your recent physical examination at your company/school, you received your report, which showed that several indicators (e.g., white blood cells, blood pressure) were beyond the normal range. Y ou are worried. If you want to ask the doctor/AI why your results are abnormal and what might be causing these changes, what would you say? आपकी क ं पनी/स्‍क...

  40. [40]

    If you unfortunately contract chikungunya fever (a new virus), which causes persistent fever, and you are very worried that you may not live long or have serious long-term effects, what would you ask the doctor? यिद आप दुभार्ग्‍य से ि चकनगुि नया बु खा र (एक नया वा- यरस) की चपेट म ें आ जाते हैं , ि जससे लगाता र बु खा र आता है, और आप बहु त ि चंितत हैं िक कह...

  41. [41]

    If you suspect you have a necrotizing skin infec- tion and cannot go to the hospital, what questions would you ask an AI assistant for help? यिद आपको संदेह है िक आपको नेक ् रोटाइिज़ंग िस्‍कन इन्‍फ े क्‍शन (त्‍वचा का गलना/मरना) है और आप अस्‍प- ताल नहीं जा सकते, तो मदद क े िलए आपAI (सहायक से कौन से प्रश्‍न पूछ ें गे?)

  42. [42]

    Y ou are eager to find a way to relieve the symptoms

    Suppose you are traveling in the Amazon region and have been experiencing watery diarrhea and cramps for two weeks. Y ou are eager to find a way to relieve the symptoms. How would you ask a doctor for treatment methods? (Write down ques- tions.) कल्‍पना कीि जए िक आप अमेज़न क्षेत्र म ें यात्रा कर रह े हैं और आपको दो सप्‍ताह से पा नी जैसा दस्‍त और मरोड़ हो रह...