Pith. sign in

REVIEW 2 major objections 4 minor 35 references

NDAI-NeuroMAP: A Neuroscience-Specific Embedding Model for Domain-Specific Retrieval

T0 review · 2 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 The paper claims that fine-tuning a 110M-parameter biomedical embedding model on 500,000 neuroscience triplets plus definitions and knowledge-graph statements yields a retrieval model with Recall@1 of 0.945 on about 24,000 held-out…

desk verdict A plausible fine-tuning recipe undermined by an in-distribution test set and no released artifacts; the headline numbers should not be trusted yet. read the letter →

arxiv 2507.03329 v1 pith:6P4LZ4Q6 submitted 2025-07-04 cs.AI

Devendra Patel , Aaditya Jain , Jayant Verma , Divyansh Rajput , Sunil Mahala , Ketki Suresh Khapare , Jayateja Kalla This is my paper · ORCID
classification cs.AI
keywords neuroscience-specificembeddingsdomain-specificretrievalcontrastivelearningknowledgedistillationretrieval-augmentedgenerationelectronichealthrecordsBioLORD-2023dense
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

The paper sets out to show that a small, domain-tuned embedding model can outperform much larger general-purpose and biomedical models for neuroscience retrieval. It describes NDAI-NeuroMAP, a 110M-parameter model fine-tuned from BioLORD-2023 on 500,000 query-definition triplets, 250,000 dictionary-style definitions, and 250,000 knowledge-graph statements. On roughly 24,000 held-out neuroscience queries, the authors report Recall@1 of 0.945 versus 0.723 for the best baseline, and a mean IoU of 0.92 versus 0.84 for BioLORD-2023 in a 100-query RAG test over neurological EHR notes. If the evaluation is a fair proxy for real retrieval, domain-specific fine-tuning at this scale would make specialized search both more accurate and far cheaper to deploy.

What carries the argument

The central machinery is a two-phase fine-tuning scheme on a sentence-transformer initialized from BioLORD-2023. Phase 1 applies a multi-objective contrastive objective: InfoNCE losses for dense, sparse, and ColBERT-style multi-vector relevance scores, fused into an ensemble score that serves as a soft teacher for self-knowledge distillation (KL divergence between the ensemble distribution and each modality). Phase 2 aligns the student to BioLORD-2023 as teacher with three losses: cosine embedding loss, MSE on embedding vectors, and Frobenius-norm loss on the batch similarity matrix. The training data are 500,000 query-definition triplets with five hard negatives, 250,000 dictionary definitions from the Apollo Corpus, and 250,000 knowledge-graph triplets from the BioLORD Dataset converted into natural-language statements. This combination is what the paper credits for the Recall@1 jump.

What would settle it

Take 500 to 1,000 neuroscience queries written by clinicians or researchers who saw none of the training data, have experts mark the relevant passages in a fixed corpus, and compare Recall@1 of NDAI-NeuroMAP against Qwen3-Embedding-4B with identical chunking and indexing. If the gap falls from the reported 22.2 points to near zero, the central claim of generalizable superiority fails.

Watch

Extended reading notes

Core claim

NDAI-NeuroMAP is presented as the first neuroscience-specific dense embedding model: starting from the BioLORD-2023 biomedical checkpoint, the authors train in two phases. Phase 1 uses 500,000 (query, positive, five negatives) triplets with InfoNCE contrastive loss over dense, sparse, and ColBERT-style scores plus self-knowledge distillation across those modalities. Phase 2 distills the original BioLORD-2023 teacher into the student on 250,000 definitions and 250,000 natural-language knowledge-graph statements using cosine, MSE, and similarity-matrix losses. The reported result is Recall@1 of 0.945, Recall@3 of 0.991, Recall@5 of 0.998, and MRR of 0.968 on about 24,000 held-out neuroscience queries, with the closest baseline (Qwen3-Embedding-4B) at 0.723, 0.901, 0.971, and 0.822. The authors interpret these numbers as evidence that neuroscience has semantic structure that neither general nor broadly biomedical embeddings capture, and that a compact domain model can deliver it at 0.42 GB GPU memory and 2,847 sequences per second.

Load-bearing premise

The load-bearing premise is that the roughly 24,000-query held-out set, built with the same LLM generation pipeline and in-house EHR notes that produced the training triplets, measures real neuroscience retrieval rather than memorized patterns.

Editorial extensions

If this is right

  • A 110M-parameter embedding can outperform models roughly 36 times larger on neuroscience retrieval, so high-accuracy domain search becomes feasible on a single GPU or even CPU.
  • RAG systems over neurological EHR notes receive better-aligned context: the paper's 100-query test shows mean IoU rising from 0.84 for BioLORD-2023 to 0.92.
  • Hybrid retrieval is available without extra infrastructure because the model emits dense, sparse, and multi-vector representations from one encoder.
  • The same two-phase data recipe can be transplanted to other scientific subdomains that suffer from terminological gaps in general embeddings.

Reading between the lines

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

  • Inference: because the held-out queries are generated by the same LLM pipeline and in-house EHR notes that produced the training triplets, the 22.2-point Recall@1 gap may partly reflect matching a synthetic style rather than generalizable neuroscience understanding; an independently written, expert-labelled test set would separate those.
  • Inference: since BioLORD-2023 supplies both the initialization and the distillation teacher, NDAI-NeuroMAP's ceiling is set by that teacher; a stronger biomedical teacher could yield further gains, and a non-biomedical teacher would likely reduce them.
  • Inference: the method's gains may come less from 'neuroscience' per se than from closing the gap between training distribution and evaluation distribution; applying the same recipe to oncology or genetics would show whether the effect is domain-specific or generic to any focused fine-tuning.
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

2 major / 4 minor

Summary. The paper presents NDAI-NeuroMAP, a 110M-parameter sentence-transformer fine-tuned from BioLORD-2023 on 500,000 LLM-generated query-definition triplets from in-house neurological EHR notes, plus 250,000 definitions and 250,000 knowledge-graph triplets. The model uses a hybrid dense/sparse/ColBERT scoring with multi-objective contrastive learning and knowledge distillation. The authors report Recall@1 of 0.945 on ~24,000 held-out neuroscience queries, beating Qwen3-Embedding-4B by 22.2 points, and a mean IoU of 0.92 versus 0.84 for BioLORD-2023 in a 100-query RAG test.

Significance. The problem is well motivated: domain-specific retrieval in neuroscience is practically important, and a compact model that genuinely outperforms larger general and biomedical models would be a useful contribution. The training recipe is detailed and the efficiency numbers (2,847 sequences/s, 0.42 GB memory) are attractive if the retrieval gains are real. However, the significance is conditional. The evaluation does not currently establish generalization: the held-out test set is not documented as independent from the training-generation pipeline, no external or human-labeled benchmark is used, and no code, data, or weights are released. These omissions are load-bearing for the headline 'substantially superior recall' claim.

major comments (2)
  1. [V.A / III.A] The ~24,000-query held-out set is described as 'carefully curated' but its construction is not specified. The only triplet-generation pipeline described in the paper (Section III.A) uses the same in-house EHR progress notes and the same LLM-assisted synthesis that produced the 500,000 training triplets. If the held-out set was generated in the same way, the 22.2-point Recall@1 gain over Qwen3-Embedding-4B in Table III may reflect in-distribution style, vocabulary, and answer-pattern overlap rather than generalizable neuroscience retrieval. In addition, each test instance has only six candidates, and the sparse scoring term in Eq. (2) rewards token overlap; a model that learns to match query terms to definitional phrases containing those terms (as in Table I) could inflate Recall@1 without semantic understanding. Please document exactly how the test set was built, demonstrate that it is non-overlapping with training data in both text and provenance, and add at least one external or human-labeled evaluation (e.g., relevance judgments over neuroscience abstracts or a clinical retrieval set) before claiming general superiority.
  2. [VIII / Appendix A] The evaluation is currently unfalsifiable: the paper states that the evaluation dataset, training code, and model weights are planned for release only upon acceptance, and the only downstream RAG evaluation in Appendix A is conducted on the same in-house EHR corpus that grounded training triplet generation. This prevents independent verification and makes the central claim difficult to assess. Please release the evaluation data and model, or provide a datasheet plus an external evaluation that does not depend on the training data source.
minor comments (4)
  1. [Table III] The Accuracy and Recall@1 columns are identical for every model; either define the distinction or remove one of the columns, since with one positive per query both metrics coincide.
  2. [IV.D.4] The text says 'distributed training setup across single NVIDIA A100 GPUs' and later 'approximately 20 hours on a single NVIDIA A100 GPU'; please clarify whether training used one GPU or multiple GPUs and state the batch size.
  3. [VI.E] The RAG evaluation is described as 'manual evaluation, with a large language model (LLM) serving as a judge' and later as 'manual testing with an LLM judge,' while Appendix A describes manual annotation of answer categories; please describe the evaluation protocol consistently and specify whether the LLM judge or human annotators produced the ground truth.
  4. [V.B] The stated criterion is models with fewer than 4 billion parameters, but the baseline list includes Qwen3-Embedding-4B; please reconcile the criterion or rename the model family.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline gain is measured on a held-out set generated by the same LLM/EHR pipeline as training; no independent benchmark is provided, so the central 'prediction' is in-distribution rather than demonstrated generalization.

  1. fitted input called prediction [Section III.A (data construction) and Section V.A (evaluation methodology), with the headline result in Table III]
    "To construct these 500,000 triplets, we leveraged our in-house patient Electronic Health Record (EHR) data, comprising progress notes exclusively from neurological patients. Using these progress notes as contextual grounding, we employed a large language model (LLM) to synthesize diverse and realistic neuroscience queries and their corresponding definitions. ... The evaluation is conducted on an expanded held-out test set comprising approximately 24,000 carefully curated neuroscience query triplets ... Each test instance follows the established triplet format (query, positive, negatives)."

    Only one procedure for constructing (query, positive, negative) triplets is described anywhere in the paper: the Section III.A pipeline that uses an LLM and in-house neurological EHR notes. The 24,000-query test set is introduced with no independent provenance and is explicitly said to follow the same triplet format. The model is fine-tuned on 500,000 triplets from that same pipeline, so the Table III claim of 0.945 vs 0.723 Recall@1 is measured on data drawn from the same synthetic distribution the model was trained to fit. Appendix A reinforces this by evaluating RAG on the same in-house patient EHR data used for training.

full rationale

The paper's central derivation is a fine-tuning recipe: initialize from BioLORD-2023, train on 500k triplets plus 250k definitions and 250k knowledge-graph statements, then report retrieval metrics on a held-out set. The training objective itself is not circular: the contrastive losses use explicit positive/negative labels, and the knowledge-distillation losses use an external teacher (BioLORD-2023), so the model's parameters are not defined in terms of the test labels. I also do not count the self-knowledge distillation (Eqs. 4-9) as circular, because the ensemble score is a combination of the model's own retrieval modalities and is standard practice. The significant circularity concern is the evaluation. The paper describes only one way to make neuroscience triplets, and that is the LLM-from-EHR procedure used for the training data. The held-out test set is said to be 'carefully curated' and to use the 'established triplet format,' but its construction is never independently described or sourced. Therefore the 22.2-point Recall@1 advantage over Qwen3-Embedding-4B is measured on the same synthetic distribution the model was trained on. This makes the headline result partly a measure of in-distribution fit to the training-generation pipeline, not a demonstrated prediction on independent, externally validated neuroscience retrieval. The lack of released data, code, or any human-labeled/external benchmark in the paper means this evaluation independence assumption is load-bearing and currently unverified. For these reasons I assign a partial circularity score of 6: the core quantitative claim reduces, to a substantial degree, to performance on the same data-generation procedure that produced the training inputs, though it is not a pure definitional equivalence.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several unverified premises: the synthetic corpora are treated as ground truth without a reported quality audit; the held-out test set is assumed to be an independent proxy for real retrieval; the hybrid scoring weights and loss weights are chosen empirically; and the teacher distillation is assumed to preserve useful biomedical knowledge. None of these premises is supported by released artifacts or external benchmarks.

free parameters (4)
  • Ensemble fusion weights w1, w2, w3 = w1=1.0, w2=0.3, w3=1.0
    Eq. (4) states these are empirically tuned weights balancing dense, sparse, and ColBERT scores; they are set by validation performance, not derived.
  • Primary loss weights lambda1, lambda2, lambda3 = lambda1=1.0, lambda2=0.1, lambda3=1.0
    Eq. (6) sets weights for dense, sparse, and ColBERT contrastive losses; the paper says less emphasis is placed on the sparse pathway early in training.
  • Distillation loss weights alpha1, alpha2, alpha3 = alpha1=alpha2=alpha3=1
    Eq. (13) sets equal weights for cosine, MSE, and similarity losses; the text says these are adjustable via validation performance.
  • Learning rate and label smoothing = 2e-5, 0.1
    Section IV.D.3 reports a learning rate of 2e-5 with cosine annealing and label smoothing of 0.1, described as empirically determined on validation sets.
assumptions (5)
  • domain assumption LLM-generated query-definition triplets and LLM-filtered ontology entries are semantically correct and representative of real neuroscience retrieval needs.
    Sections III.A-C describe using an LLM to synthesize queries, definitions, and KG statements from EHR notes and public corpora; no human verification or quality audit is reported.
  • domain assumption The held-out test set is independent of the training set and representative of real-world retrieval.
    Section V.A and Section VIII: no leakage analysis or external benchmark is reported, and the test set is curated from the same data sources and generation style as the training set.
  • domain assumption Dense, sparse, and ColBERT scores can be linearly combined into a single relevance score with fixed weights.
    Eq. (4) fuses scores with fixed weights; no theoretical justification, ablation, or sensitivity analysis is reported.
  • domain assumption Cosine and MSE distillation toward BioLORD-2023 preserves useful biomedical knowledge while allowing neuroscience specialization.
    Section IV.C relies on the teacher's quality and on hand-set loss weights; no analysis of forgetting or of the teacher-student gap is provided.
  • standard math InfoNCE and KL-divergence self-distillation are valid objectives for this fine-tuning setup.
    These are standard loss functions in embedding training and are not themselves controversial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NDAI-NeuroMAP: A Neuroscience-Specific Embedding Model for Domain-Specific Retrieval." pith.science (2026). https://pith.science/paper/6P4LZ4Q6

@misc{pith2026250703329,
  author       = {Pith},
  title        = {Pith review of: NDAI-NeuroMAP: A Neuroscience-Specific Embedding Model for Domain-Specific Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6P4LZ4Q6}},
  note         = {Machine review of arXiv:2507.03329}
}
read the original abstract

We present NDAI-NeuroMAP, the first neuroscience-domain-specific dense vector embedding model engineered for high-precision information retrieval tasks. Our methodology encompasses the curation of an extensive domain-specific training corpus comprising 500,000 carefully constructed triplets (query-positive-negative configurations), augmented with 250,000 neuroscience-specific definitional entries and 250,000 structured knowledge-graph triplets derived from authoritative neurological ontologies. We employ a sophisticated fine-tuning approach utilizing the FremyCompany/BioLORD-2023 foundation model, implementing a multi-objective optimization framework combining contrastive learning with triplet-based metric learning paradigms. Comprehensive evaluation on a held-out test dataset comprising approximately 24,000 neuroscience-specific queries demonstrates substantial performance improvements over state-of-the-art general-purpose and biomedical embedding models. These empirical findings underscore the critical importance of domain-specific embedding architectures for neuroscience-oriented RAG systems and related clinical natural language processing applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 21 canonical work pages

  1. [13]

    Biolord-2023: A biomedical language representation model

    Guillaume Fremy, Marie Dubois, Pierre Martin, and Sophie Laurent. Biolord-2023: A biomedical language representation model. Journal of Biomedical Semantics, 14(1):1–15, 2023. 12

    work page 2023

  2. [29]

    Seeing the Fruit for the Leaves: Robotically Mapping Apple Fruitlets in a Commercial Orchard

    Ewan Dunbar, Jianmo Chen, and Luke Zhang. Stella: Efficient neural text embeddings for retrieval. arXiv preprint arXiv:2308.07512 , 2023

    work page Pith review arXiv 2023

  3. [19]

    Gate-controlled neuromorphic functional transition in an electrochemical graphene transistor

    Xuansheng Li, Kexin Wang, and Danqi Chen. Jasper and stella: Distillation of sota embedding models. arXiv preprint arXiv:2312.04934 , 2023

    work page Pith review arXiv 2023

  4. [1]

    Efficient estimation of word representations in vector space

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013

    arXiv 2013

  5. [2]

    Glove: Global vectors for word representation

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 1532–1543, 2014

    work page 2014

  6. [3]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 , 2018

    arXiv 2018

  7. [4]

    Roberta: A robustly optimized bert pretraining approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 , 2019

    arXiv 1907

  8. [5]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019

    arXiv 1910

Show all 35 references
  1. [6]

    Text and code embeddings by contrastive pre-training

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005 , 2022

    2022 arXiv

  2. [7]

    Towards general text embeddings with multi-stage contrastive learning

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 , 2023

    2023 arXiv

  3. [8]

    Evaluating the performance of general-purpose language models on domain-specific tasks

    Buzhou Tang, Ying Qin, Xiaolong Liu, Tao Wang, Ruifeng Li, and Jianbo Lei. Evaluating the performance of general-purpose language models on domain-specific tasks. Journal of Biomedical Informatics , 103:103378, 2020

    2020

  4. [9]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240, 2020

    2020

  5. [10]

    Publicly available clinical bert embeddings

    Emily Alsentzer, John R Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings. arXiv preprint arXiv:1904.03323 , 2019

    1904 arXiv

  6. [11]

    The unified medical language system (umls): integrating biomedical terminology

    Olivier Bodenreider. The unified medical language system (umls): integrating biomedical terminology. Nucleic acids research, 32(suppl 1):D267–D270, 2004

    2004

  7. [12]

    Bibliometric analysis on the literature of neuroscience

    Andy Wai Kan Yeung, Tetsuo K Goto, and W Keung Leung. Bibliometric analysis on the literature of neuroscience. Frontiers in Neuroscience, 11:718, 2017

    2017

  8. [14]

    Scibert: A pretrained language model for scientific text

    Iz Beltagy, Kyle Lo, and Arman Cohan. Scibert: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , ...

    2019

  9. [15]

    Enhancing clinical decision support with retrieval-augmented generation

    Aditya Patel, John Smith, Mary Johnson, and David Brown. Enhancing clinical decision support with retrieval-augmented generation. In Proceedings of the 2023 Conference on Health Informatics , pages 123–130, 2023

    2023

  10. [16]

    Apollocorpus

    Freedom Intelligence. Apollocorpus. https://huggingface.co/datasets/FreedomIntelligence/ApolloCorpus, 2023. Accessed: 2025-07-01

    2023

  11. [17]

    Biolord dataset

    Fremy Company. Biolord dataset. https://huggingface.co/datasets/FremyCompany/BioLORD-Dataset, 2023. Accessed: 2025-07-01

    2023

  12. [18]

    M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 , 2024

    2024 arXiv

  13. [20]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  14. [21]

    Curriculum learning

    Yoshua Bengio, J ´erˆome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. Proceedings of the 26th annual international conference on machine learning , pages 41–48, 2009

    2009

  15. [22]

    Sgdr: Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations , 2017

    2017

  16. [23]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 2818–2826, 2016

    2016

  17. [24]

    Nccl: Optimized primitives for collective multi-gpu communication

    NVIDIA Corporation. Nccl: Optimized primitives for collective multi-gpu communication. https://developer.nvidia.com/nccl, 2017

    2017

  18. [25]

    Training deep nets with sublinear memory cost

    Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174 , 2016

    2016 arXiv

  19. [26]

    Mixed precision training

    Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740 , 2017

    arXiv

  20. [27]

    Practical bayesian optimization of machine learning algorithms

    Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, volume 25, 2012

    2012

  21. [28]

    Biolord-2023 model

    FremyCompany. Biolord-2023 model. https://huggingface.co/FremyCompany/BioLORD-2023, 2023

    2023

  22. [30]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  23. [31]

    Statistical Power Analysis for the Behavioral Sciences

    Jacob Cohen. Statistical Power Analysis for the Behavioral Sciences . Lawrence Erlbaum Associates, Hillsdale, NJ, 2nd edition, 1988

    1988

  24. [32]

    Sittig, and Adam Wright

    Blackford Middleton, Dean F. Sittig, and Adam Wright. Clinical decision support systems for the practice of evidence-based medicine. Journal of the American Medical Informatics Association , 23(6):1057–1065, 2016

    2016

  25. [33]

    Bright, Anthony Wong, Radhika Dhurjati, Erin Bristow, Lori Bastian, Remy R

    Timothy J. Bright, Anthony Wong, Radhika Dhurjati, Erin Bristow, Lori Bastian, Remy R. Coeytaux, Gregory Samsa, Vic Hasselblad, John W. Williams, Michael D. Musty, et al. Effect of clinical decision-support systems: A systematic review. Annals of Internal Medicine , 157(1):29–43, 2012

    2012

  26. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. International Conference on Machine Learning , page...

    2021

  27. [35]

    What medications was the patient taking at the last visit?

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of scien...

    2017

Pith tools

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