Tathagata Raha

Clément Christophe, Praveen K Kanithi, Tathagata Raha et al.

Med42-v2 introduces a suite of clinical large language models (LLMs) designed to address the limitations of generic models in healthcare settings. These models are built on Llama3 architecture and fine-tuned using specialized clinical data. They underwent multi-stage preference alignment to effectively respond to natural prompts. While generic models are often preference-aligned to avoid answering clinical queries as a precaution, Med42-v2 is specifically trained to overcome this limitation, enabling its use in clinical settings. Med42-v2 models demonstrate superior performance compared to the original Llama3 models in both 8B and 70B parameter configurations and GPT-4 across various medical benchmarks. These LLMs are developed to understand clinical queries, perform reasoning tasks, and provide valuable assistance in clinical environments. The models are now publicly available at \href{https://huggingface.co/m42-health}{https://huggingface.co/m42-health}.

Praveenkumar Kanithi, Clément Christophe, Marco AF Pimentel et al.

While Large Language Models (LLMs) achieve superhuman performance on standardized medical licensing exams, these static benchmarks have become saturated and increasingly disconnected from the functional requirements of clinical workflows. To bridge the gap between theoretical capability and verified utility, we introduce MEDIC, a comprehensive evaluation framework establishing leading indicators across various clinical dimensions. Beyond standard question-answering, we assess operational capabilities using deterministic execution protocols and a novel Cross-Examination Framework (CEF), which quantifies information fidelity and hallucination rates without reliance on reference texts. Our evaluation across a heterogeneous task suite exposes critical performance trade-offs: we identify a significant knowledge-execution gap, where proficiency in static retrieval does not predict success in operational tasks such as clinical calculation or SQL generation. Furthermore, we observe a divergence between passive safety (refusal) and active safety (error detection), revealing that models fine-tuned for high refusal rates often fail to reliably audit clinical documentation for factual accuracy. These findings demonstrate that no single architecture dominates across all dimensions, highlighting the necessity of a portfolio approach to clinical model deployment. As part of this investigation, we released a public leaderboard on Hugging Face.\footnote{https://huggingface.co/spaces/m42-health/MEDIC-Benchmark}

Tathagata Raha, Mukund Choudhary, Abhinav Menon et al.

Factual consistency is one of the most important requirements when editing high quality documents. It is extremely important for automatic text generation systems like summarization, question answering, dialog modeling, and language modeling. Still, automated factual inconsistency detection is rather under-studied. Existing work has focused on (a) finding fake news keeping a knowledge base in context, or (b) detecting broad contradiction (as part of natural language inference literature). However, there has been no work on detecting and explaining types of factual inconsistencies in text, without any knowledge base in context. In this paper, we leverage existing work in linguistics to formally define five types of factual inconsistencies. Based on this categorization, we contribute a novel dataset, FICLE (Factual Inconsistency CLassification with Explanation), with ~8K samples where each sample consists of two sentences (claim and context) annotated with type and span of inconsistency. When the inconsistency relates to an entity type, it is labeled as well at two levels (coarse and fine-grained). Further, we leverage this dataset to train a pipeline of four neural models to predict inconsistency type with explanations, given a (claim, context) sentence pair. Explanations include inconsistent claim fact triple, inconsistent context span, inconsistent claim component, coarse and fine-grained inconsistent entity types. The proposed system first predicts inconsistent spans from claim and context; and then uses them to predict inconsistency types and inconsistent entity types (when inconsistency is due to entities). We experiment with multiple Transformer-based natural language classification as well as generative models, and find that DeBERTa performs the best. Our proposed methods provide a weighted F1 of ~87% for inconsistency type classification across the five classes.

Marco AF Pimentel, Clément Christophe, Tathagata Raha et al.

As large language models (LLMs) continue to evolve, the need for robust and standardized evaluation benchmarks becomes paramount. Evaluating the performance of these models is a complex challenge that requires careful consideration of various linguistic tasks, model architectures, and benchmarking methodologies. In recent years, various frameworks have emerged as noteworthy contributions to the field, offering comprehensive evaluation tests and benchmarks for assessing the capabilities of LLMs across diverse domains. This paper provides an exploration and critical analysis of some of these evaluation methodologies, shedding light on their strengths, limitations, and impact on advancing the state-of-the-art in natural language processing.

Clément Christophe, Tathagata Raha, Svetlana Maslenkova et al.

Large Language Models (LLMs) have demonstrated significant potential in transforming clinical applications. In this study, we investigate the efficacy of four techniques in adapting LLMs for clinical use-cases: continuous pretraining, instruct fine-tuning, NEFTune, and prompt engineering. We employ these methods on Mistral 7B and Mixtral 8x7B models, leveraging a large-scale clinical pretraining dataset of 50 billion tokens and an instruct fine-tuning dataset of 500 million tokens. Our evaluation across various clinical tasks reveals the impact of each technique. While continuous pretraining beyond 250 billion tokens yields marginal improvements on its own, it establishes a strong foundation for instruct fine-tuning. Notably, NEFTune, designed primarily to enhance generation quality, surprisingly demonstrates additional gains on our benchmark. Complex prompt engineering methods further enhance performance. These findings show the importance of tailoring fine-tuning strategies and exploring innovative techniques to optimize LLM performance in the clinical domain.

Tathagata Raha, Clement Christophe, Nada Saadi et al.

Traditional metrics like BLEU and BERTScore fail to capture semantic fidelity in generative text-to-text tasks. We adapt the Cross-Examination Framework (CEF) for a reference-free, multi-dimensional evaluation by treating the source and candidate as independent knowledge bases. CEF generates verifiable questions from each text and performs a cross-examination to derive three interpretable scores: Coverage, Conformity, and Consistency. Validated across translation, summarization and clinical note-generation, our framework identifies critical errors, such as content omissions and factual contradictions, missed by standard metrics. A key contribution is a systematic robustness analysis to select a stable judge model. Crucially, the strong correlation between our reference-free and with-reference modes validates CEF's reliability without gold references. Furthermore, human expert validation demonstrates that CEF mismatching questions align with meaning-altering semantic errors higher than with non-semantic errors, particularly excelling at identifying entity-based and relational distortions.

Clément Christophe, Wadood Mohammed Abdul, Prateek Munjal et al.

As LLMs are increasingly integrated into clinical workflows, their tendency for sycophancy, prioritizing user agreement over factual accuracy, poses significant risks to patient safety. While existing evaluations often rely on subjective datasets, we introduce a robust framework grounded in medical MCQA with verifiable ground truths. We propose the Adjusted Sycophancy Score, a novel metric that isolates alignment bias by accounting for stochastic model instability, or "confusability". Through an extensive scaling analysis of the Qwen-3 and Llama-3 families, we identify a clear scaling trajectory for resilience. Furthermore, we reveal a counter-intuitive vulnerability in reasoning-optimized "Thinking" models: while they demonstrate high vanilla accuracy, their internal reasoning traces frequently rationalize incorrect user suggestions under authoritative pressure. Our results across frontier models suggest that benchmark performance is not a proxy for clinical reliability, and that simplified reasoning structures may offer superior robustness against expert-driven sycophancy.

Clément Christophe, Praveen K Kanithi, Prateek Munjal et al.

This study presents a comprehensive analysis and comparison of two predominant fine-tuning methodologies - full-parameter fine-tuning and parameter-efficient tuning - within the context of medical Large Language Models (LLMs). We developed and refined a series of LLMs, based on the Llama-2 architecture, specifically designed to enhance medical knowledge retrieval, reasoning, and question-answering capabilities. Our experiments systematically evaluate the effectiveness of these tuning strategies across various well-known medical benchmarks. Notably, our medical LLM Med42 showed an accuracy level of 72% on the US Medical Licensing Examination (USMLE) datasets, setting a new standard in performance for openly available medical LLMs. Through this comparative analysis, we aim to identify the most effective and efficient method for fine-tuning LLMs in the medical domain, thereby contributing significantly to the advancement of AI-driven healthcare applications.

Nada Saadi, Tathagata Raha, Clément Christophe et al.

This paper investigates the challenges of developing large language models (LLMs) proficient in both multilingual understanding and medical knowledge. We demonstrate that simply translating medical data does not guarantee strong performance on clinical tasks in the target language. Our experiments reveal that the optimal language mix in training data varies significantly across different medical tasks. We find that larger models with carefully calibrated language ratios achieve superior performance on native-language clinical tasks. Furthermore, our results suggest that relying solely on fine-tuning may not be the most effective approach for incorporating new language knowledge into LLMs. Instead, data and computationally intensive pretraining methods may still be necessary to achieve optimal performance in multilingual medical settings. These findings provide valuable guidance for building effective and inclusive medical AI systems for diverse linguistic communities.

Svetlana Maslenkova, Clement Christophe, Marco AF Pimentel et al.

Large language models offer transformative potential for healthcare, yet their responsible and equitable development depends critically on a deeper understanding of how training data characteristics influence model behavior, including the potential for bias. Current practices in dataset curation and bias assessment often lack the necessary transparency, creating an urgent need for comprehensive evaluation frameworks to foster trust and guide improvements. In this study, we present an in-depth analysis of potential downstream biases in clinical language models, with a focus on differential opioid prescription tendencies across diverse demographic groups, such as ethnicity, gender, and age. As part of this investigation, we introduce HC4: Healthcare Comprehensive Commons Corpus, a novel and extensively curated pretraining dataset exceeding 89 billion tokens. Our evaluation leverages both established general benchmarks and a novel, healthcare-specific methodology, offering crucial insights to support fairness and safety in clinical AI applications.

Tathagata Raha, Vijayasaradhi Indurthi, Aayush Upadhyaya et al.

The evolution of social media platforms have empowered everyone to access information easily. Social media users can easily share information with the rest of the world. This may sometimes encourage spread of fake news, which can result in undesirable consequences. In this work, we train models which can identify health news related to COVID-19 pandemic as real or fake. Our models achieve a high F1-score of 98.64%. Our models achieve second place on the leaderboard, tailing the first position with a very narrow margin 0.05% points.

Tathagata Raha, Sayar Ghosh Roy, Ujwal Narayan et al.

Identifying adverse and hostile content on the web and more particularly, on social media, has become a problem of paramount interest in recent years. With their ever increasing popularity, fine-tuning of pretrained Transformer-based encoder models with a classifier head are gradually becoming the new baseline for natural language classification tasks. In our work, we explore the gains attributed to Task Adaptive Pretraining (TAPT) prior to fine-tuning of Transformer-based architectures. We specifically study two problems, namely, (a) Coarse binary classification of Hindi Tweets into Hostile or Not, and (b) Fine-grained multi-label classification of Tweets into four categories: hate, fake, offensive, and defamation. Building up on an architecture which takes emojis and segmented hashtags into consideration for classification, we are able to experimentally showcase the performance upgrades due to TAPT. Our system (with team name 'iREL IIIT') ranked first in the 'Hostile Post Detection in Hindi' shared task with an F1 score of 97.16% for coarse-grained detection and a weighted F1 score of 62.96% for fine-grained multi-label classification on the provided blind test corpora.

Sayar Ghosh Roy, Ujwal Narayan, Tathagata Raha et al.

Detecting and classifying instances of hate in social media text has been a problem of interest in Natural Language Processing in the recent years. Our work leverages state of the art Transformer language models to identify hate speech in a multilingual setting. Capturing the intent of a post or a comment on social media involves careful evaluation of the language style, semantic content and additional pointers such as hashtags and emojis. In this paper, we look at the problem of identifying whether a Twitter post is hateful and offensive or not. We further discriminate the detected toxic content into one of the following three classes: (a) Hate Speech (HATE), (b) Offensive (OFFN) and (c) Profane (PRFN). With a pre-trained multilingual Transformer-based text encoder at the base, we are able to successfully identify and classify hate speech from multiple languages. On the provided testing corpora, we achieve Macro F1 scores of 90.29, 81.87 and 75.40 for English, German and Hindi respectively while performing hate speech detection and of 60.70, 53.28 and 49.74 during fine-grained classification. In our experiments, we show the efficacy of Perspective API features for hate speech classification and the effects of exploiting a multilingual training scheme. A feature selection study is provided to illustrate impacts of specific features upon the architecture's classification head.

Tathagata Raha, Sainik Kumar Mahata, Dipankar Das et al.

Code-mixed texts are widespread nowadays due to the advent of social media. Since these texts combine two languages to formulate a sentence, it gives rise to various research problems related to Natural Language Processing. In this paper, we try to excavate one such problem, namely, Parts of Speech tagging of code-mixed texts. We have built a system that can POS tag English-Bengali code-mixed data where the Bengali words were written in Roman script. Our approach initially involves the collection and cleaning of English-Bengali code-mixed tweets. These tweets were used as a development dataset for building our system. The proposed system is a modular approach that starts by tagging individual tokens with their respective languages and then passes them to different POS taggers, designed for different languages (English and Bengali, in our case). Tags given by the two systems are later joined together and the final result is then mapped to a universal POS tag set. Our system was checked using 100 manually POS tagged code-mixed sentences and it returned an accuracy of 75.29%