|
--- |
|
license: mit |
|
datasets: |
|
- disi-unibo-nlp/medqa-5-opt-MedGENIE |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
tags: |
|
- medical |
|
- question-answering |
|
- fusion-in-decoder |
|
pipeline_tag: question-answering |
|
widget: |
|
- text: >- |
|
A junior orthopaedic surgery resident is completing a carpal tunnel repair |
|
with the department chairman as the attending physician. During the case, |
|
the resident inadvertently cuts a flexor tendon. The tendon is repaired |
|
without complication. The attending tells the resident that the patient will |
|
do fine, and there is no need to report this minor complication that will |
|
not harm the patient, as he does not want to make the patient worry |
|
unnecessarily. He tells the resident to leave this complication out of the |
|
operative report. Which of the following is the correct next action for the |
|
resident to take? A. Disclose the error to the patient and put it in the |
|
operative report B. Tell the attending that he cannot fail to disclose this |
|
mistake C. Report the physician to the ethics committee D. Refuse to dictate |
|
the operative reporty. |
|
context: >- |
|
Inadvertent Cutting of Tendon is a complication, it should be in the |
|
Operative Reports The resident must put this complication in the operative |
|
report and disscuss it with the patient. If there was no harm to the patent |
|
and correction was done then theres nothing major for worry. But disclosing |
|
this as per ethical guidelines, is mandatory |
|
example_title: Example 1 |
|
--- |
|
# Model Card for MedGENIE-fid-flan-t5-base-medqa |
|
|
|
MedGENIE comprises a collection of language models designed to utilize generated contexts, rather than retrieved ones, for addressing multiple-choice open-domain questions in the medical field. Specifically, **MedGENIE-fid-flan-t5-base-medqa** is a *fusion-in-decoder* (FID) model based on [flan-t5-base](https://huggingface.co/google/flan-t5-base), trained on the [MedQA-USMLE](https://huggingface.co/datasets/disi-unibo-nlp/medqa-5-opt-MedGENIE) dataset and grounded on artificial contexts generated by [PMC-LLaMA-13B](https://huggingface.co/axiong/PMC_LLaMA_13B). This model achieves a new *state-of-the-art* (SOTA) performance over the corresponding test set. |
|
|
|
## Model description |
|
|
|
- **Language(s) (NLP):** English |
|
- **License:** MIT |
|
- **Finetuned from model:** [google/flan-t5-base](https://huggingface.co/google/flan-t5-base) |
|
- **Repository:** https://github.com/disi-unibo-nlp/medgenie |
|
- **Paper:** [To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering](https://arxiv.org/abs/2403.01924) |
|
|
|
## Performance |
|
|
|
At the time of release (February 2024), **MedGENIE-fid-flan-t5-base-medqa** is a new lightweight SOTA model on MedQA-USMLE benchmark: |
|
|
|
| Model | Ground (Source) | Learning | Params | Accuracy (↓) | |
|
|----------------------------------|--------------------|---------------------------|-----------------|-------------------------------| |
|
| **MedGENIE-FID-Flan-T5** | **G (PMC-LLaMA)** | **Fine-tuned** | **250M** | **53.1** | |
|
| Codex <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-zhot | 175B | 52.5 | |
|
| Codex <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | R (Wikipedia) | 0-shot | 175B | 52.5 | |
|
| GPT-3.5-Turbo <small>([Yang et al.](https://arxiv.org/abs/2309.02233))</small> | R (Wikipedia) | k-shot | -- | 52.3 | |
|
| MEDITRON <small>([Chen et al.](https://arxiv.org/abs/2311.16079))</small> | ∅ | Fine-tuned | 7B | 52.0 | |
|
| BioMistral DARE <small> ([Labrak et al.](https://arxiv.org/abs/2402.10373)) </small> | ∅ | Fine-tuned | 7B | 51.1 | |
|
| BioMistral <small> ([Labrak et al.](https://arxiv.org/abs/2402.10373)) </small> | ∅ | Fine-tuned | 7B | 50.6 | |
|
| Zephyr-β | R (MedWiki) | 2-shot | 7B | 50.4 | |
|
| BioMedGPT <small>([Luo et al.](https://arxiv.org/abs/2308.09442v2))</small> | ∅ | k-shot | 10B | 50.4 | |
|
| BioMedLM <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | Fine-tuned | 2.7B | 50.3 | |
|
| PMC-LLaMA <small>(awq 4 bit)</small> | ∅ | Fine-tuned | 13B | 50.2 | |
|
| LLaMA-2 <small>([Chen et al.](https://arxiv.org/abs/2311.16079))</small> | ∅ | Fine-tuned | 7B | 49.6 | |
|
| Zephyr-β | ∅ | 2-shot | 7B | 49.6 | |
|
| Zephyr-β <small>([Chen et al.](https://arxiv.org/abs/2311.16079))</small> | ∅ | 3-shot | 7B | 49.2 | |
|
| PMC-LLaMA <small>([Chen et al.](https://arxiv.org/abs/2311.16079))</small> | ∅ | Fine-tuned | 7B | 49.2 | |
|
| DRAGON <small>([Yasunaga et al.](https://arxiv.org/abs/2210.09338))</small> | R (UMLS) | Fine-tuned | 360M | 47.5 | |
|
| InstructGPT <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | R (Wikipedia) | 0-shot | 175B | 47.3 | |
|
| BioMistral DARE <small> ([Labrak et al.](https://arxiv.org/abs/2402.10373)) </small> | ∅ | 3-shot | 7B | 47.0 | |
|
| Flan-PaLM <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | 5-shot | 62B | 46.1 | |
|
| InstructGPT <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 175B | 46.0 | |
|
| VOD <small>([Liévin et al. 2023](https://arxiv.org/abs/2210.06345))</small> | R (MedWiki) | Fine-tuned | 220M | 45.8 | |
|
| Vicuna 1.3 <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 33B | 45.2 | |
|
| BioLinkBERT <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | Fine-tuned | 340M | 45.1 | |
|
| Mistral-Instruct | R (MedWiki) | 2-shot | 7B | 45.1 | |
|
| BioMistral <small> ([Labrak et al.](https://arxiv.org/abs/2402.10373)) </small> | ∅ | 3-shot | 7B | 44.4 | |
|
| Galactica | ∅ | 0-shot | 120B | 44.4 | |
|
| LLaMA-2 <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 70B | 43.4 | |
|
| BioReader <small>([Frisoni et al.](https://aclanthology.org/2022.emnlp-main.390/))</small> | R (PubMed-RCT) | Fine-tuned | 230M | 43.0 | |
|
| Guanaco <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 33B | 42.9 | |
|
| LLaMA-2-chat <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 70B | 42.3 | |
|
| Vicuna 1.5 <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 65B | 41.6 | |
|
| Mistral-Instruct <small>([Chen et al.](https://arxiv.org/abs/2311.16079))</small> | ∅ | 3-shot | 7B | 41.1 | |
|
| PaLM <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | 5-shot | 62B | 40.9 | |
|
| Guanaco <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 65B | 40.8 | |
|
| Falcon-Instruct <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 40B | 39.0 | |
|
| Vicuna 1.3 <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 13B | 38.7 | |
|
| GreaseLM <small>([Zhang et al.](https://arxiv.org/abs/2201.08860))</small> | R (UMLS) | Fine-tuned | 359M | 38.5 | |
|
| PubMedBERT <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | Fine-tuned | 110M | 38.1 | |
|
| QA-GNN <small>([Yasunaga et al.](https://arxiv.org/abs/2104.06378))</small> | R (UMLS) | Fine-tuned | 360M | 38.0 | |
|
| LLaMA-2 <small>([Yang et al.](https://arxiv.org/abs/2309.02233))</small> | R (Wikipedia) | k-shot | 13B | 37.6 | |
|
| LLaMA-2-chat | R (MedWiki) | 2-shot | 7B | 37.2 | |
|
| LLaMA-2-chat | ∅ | 2-shot | 7B | 37.2 | |
|
| BioBERT <small>([Lee et al.](https://arxiv.org/abs/1901.08746))</small> | ∅ | Fine-tuned | 110M | 36.7 | |
|
| MTP-Instruct <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 30B | 35.1 | |
|
| GPT-Neo <small>([Singhal et al.](https://arxiv.org/abs/2212.13138))</small> | ∅ | Fine-tuned | 2.5B | 33.3 | |
|
| LLaMa-2-chat <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 13B | 32.2 | |
|
| LLaMa-2 <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 13B | 31.1 | |
|
| GPT-NeoX <small>([Liévin et al.](https://arxiv.org/abs/2207.08143))</small> | ∅ | 0-shot | 20B | 26.9 | |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 5e-05 |
|
- n_context: 5 |
|
- per_gpu_batch_size: 1 |
|
- accumulation_steps: 4 |
|
- total_steps: 40,712 |
|
- eval_freq: 10,178 |
|
- optimizer: AdamW |
|
- scheduler: linear |
|
- weight_decay: 0.01 |
|
- warmup_ratio: 0.1 |
|
- text_maxlength: 1024 |
|
|
|
### Bias, Risk and Limitation |
|
|
|
Our model is trained on artificially generated contextual documents, which might inadvertently magnify inherent biases and depart from clinical and societal norms. This could lead to the spread of convincing medical misinformation. To mitigate this risk, we recommend a cautious approach: domain experts should manually review any output before real-world use. This ethical safeguard is crucial to prevent the dissemination of potentially erroneous or misleading information, particularly within clinical and scientific circles. |
|
|
|
## Citation |
|
|
|
If you find MedGENIE-fid-flan-t5-base-medqa is useful in your work, please cite it with: |
|
|
|
``` |
|
@misc{frisoni2024generate, |
|
title={To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering}, |
|
author={Giacomo Frisoni and Alessio Cocchieri and Alex Presepi and Gianluca Moro and Zaiqiao Meng}, |
|
year={2024}, |
|
eprint={2403.01924}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.CL} |
|
} |
|
``` |
