Overcoming Dataset Scarcity for Robust Radiology Impression Generation with Expert-Specific LLMs
This repository provides the pretrained model weights for our paper Overcoming Dataset Scarcity for Robust Radiology Impression Generation with Expert-Specific LLMs.
Model Nomenclature
| Model group | Model name | Model description |
|---|---|---|
| Benchmark SFT models | bloom_1b1_3 |
Bloomz-1B benchmark SFT model trained for 3 epochs on the 360,001-report SFT cohort. |
bloom_1b1_16 |
Bloomz-1B benchmark SFT model trained for 16 epochs on the 360,001-report SFT cohort. | |
bloom_3b_3 |
Bloomz-3B benchmark SFT model trained for 3 epochs on the 360,001-report SFT cohort. | |
bloom_3b_16 |
Bloomz-3B benchmark SFT model trained for 16 epochs on the 360,001-report SFT cohort. | |
| RLHF checkpoints | rlhf_checkpoint-40 |
RLHF model initialized from bloom_3b_3 and trained after 40 PPO steps. |
rlhf_checkpoint-80 |
RLHF model initialized from bloom_3b_3 and trained after 80 PPO steps. |
|
rlhf_checkpoint-120 |
RLHF model initialized from bloom_3b_3 and trained after 120 PPO steps. |
|
rlhf_checkpoint-160 |
RLHF model initialized from bloom_3b_3 and trained after 160 PPO steps. |
|
rlhf_checkpoint-200 |
RLHF model initialized from bloom_3b_3 and trained after 200 PPO steps. |
|
rlhf_checkpoint-240 |
RLHF model initialized from bloom_3b_3 and trained after 240 PPO steps. |
|
rlhf_checkpoint-280 |
RLHF model initialized from bloom_3b_3 and trained after 280 PPO steps. |
|
| Expert-specific models | 3b_radiologist1 |
Bloomz-3B expert-specific SFT model trained for 3 epochs on expert 1's cohort of 5,606 reports. |
3b_radiologist2 |
Bloomz-3B expert-specific SFT model trained for 3 epochs on expert 2's cohort of 9,016 reports. | |
3b_radiologist3 |
Bloomz-3B expert-specific SFT model trained for 3 epochs on expert 3's cohort of 4,375 reports. | |
3b_radiologist4 |
Bloomz-3B expert-specific SFT model trained for 3 epochs on expert 4's cohort of 5,916 reports. | |
3b_radiologist5 |
Bloomz-3B expert-specific SFT model trained for 3 epochs on expert 5's cohort of 2,175 reports. | |
7b_radiologist1 |
Bloomz-7B expert-specific SFT model trained for 3 epochs on expert 1's cohort of 5,606 reports. | |
7b_radiologist2 |
Bloomz-7B expert-specific SFT model trained for 3 epochs on expert 2's cohort of 9,016 reports. | |
7b_radiologist3 |
Bloomz-7B expert-specific SFT model trained for 3 epochs on expert 3's cohort of 4,375 reports. | |
7b_radiologist4 |
Bloomz-7B expert-specific SFT model trained for 3 epochs on expert 4's cohort of 5,916 reports. | |
7b_radiologist5 |
Bloomz-7B expert-specific SFT model trained for 3 epochs on expert 5's cohort of 2,175 reports. | |
| General-purpose public LLMs | llm_gpt35 |
gpt-3.5-turbo-0125 evaluated using fixed 3-shot prompting. |
llm_gpt4 |
gpt-4-0125-preview evaluated using fixed 3-shot prompting. | |
llm_baidu |
Qianfan-Chinese-Llama-2-13B evaluated using fixed 3-shot prompting. |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support