dejori commited on
Commit
d3cd80a
·
verified ·
1 Parent(s): a7e6400

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +70 -0
README.md ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - medical
5
+ - clinical-notes
6
+ - patient-communication
7
+ - lora
8
+ - peft
9
+ - medgemma
10
+ language:
11
+ - en
12
+ library_name: peft
13
+ ---
14
+
15
+ # NoteExplain Models
16
+
17
+ Trained models for clinical note simplification - translating medical documents into patient-friendly language.
18
+
19
+ ## Models
20
+
21
+ | Model | Base | Description | Overall | Accuracy | Patient-Centered |
22
+ |-------|------|-------------|---------|----------|------------------|
23
+ | **gemma-2b-distilled** | gemma-2-2b-it | Final mobile model | 70% | 73% | **76%** |
24
+ | **gemma-2b-dpo** | gemma-2-2b-it | DPO comparison | **73%** | **82%** | 61% |
25
+ | **gemma-9b-dpo** | gemma-2-9b-it | Teacher model | 79% | 91% | 70% |
26
+
27
+ ## Usage
28
+
29
+ ```python
30
+ from peft import PeftModel
31
+ from transformers import AutoModelForCausalLM, AutoTokenizer
32
+
33
+ # Load the distilled model (recommended for deployment)
34
+ base_model = AutoModelForCausalLM.from_pretrained("google/gemma-2-2b-it")
35
+ tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")
36
+ model = PeftModel.from_pretrained(base_model, "dejori/note-explain", subfolder="gemma-2b-distilled")
37
+
38
+ # Or load the DPO model (higher accuracy)
39
+ model = PeftModel.from_pretrained(base_model, "dejori/note-explain", subfolder="gemma-2b-dpo")
40
+
41
+ # Generate
42
+ prompt = "Simplify this clinical note for a patient:\n\n[clinical note]\n\nSimplified version:"
43
+ inputs = tokenizer(prompt, return_tensors="pt")
44
+ outputs = model.generate(**inputs, max_new_tokens=512)
45
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
46
+ ```
47
+
48
+ ## Training
49
+
50
+ - **DPO Training**: MedGemma-27B scored 5 candidate outputs per clinical note, creating preference pairs
51
+ - **Distillation**: 9B-DPO model generated high-quality outputs to train the 2B model via SFT
52
+
53
+ ## Dataset
54
+
55
+ Training data: [dejori/note-explain-clinical](https://huggingface.co/datasets/dejori/note-explain-clinical)
56
+
57
+ ## Citation
58
+
59
+ ```bibtex
60
+ @misc{noteexplain2026,
61
+ title={NoteExplain: Privacy-First Clinical Note Simplification},
62
+ author={Dejori, Mathaeus},
63
+ year={2026},
64
+ publisher={HuggingFace}
65
+ }
66
+ ```
67
+
68
+ ## License
69
+
70
+ Apache 2.0