Hugging Face's logo

Phora68
/
rapha

Safetensors
GGUF
English
qwen2
clinical
medical
healthcare
qlora
unsloth
chatml
rapha
8-bit precision
conversational
Model card Files
xet
 Community
Phora68 commited on
Commit
0d28b84
·
verified ·
1 Parent(s): 47cb96b

Update model card

Browse files
Files changed (1) hide show
  1. README.md +16 -16
README.md CHANGED
@@ -2,15 +2,15 @@
2
  license: other
3
  base_model: unsloth/Qwen2.5-3B-Instruct-bnb-4bit
4
  tags:
5
- - clinical
6
- - medical
7
- - healthcare
8
- - qlora
9
- - unsloth
10
- - chatml
11
- - rapha
12
  language:
13
- - en
14
  ---
15
 
16
  # Rapha — Clinical AI Physician Assistant
@@ -24,7 +24,7 @@ never diagnoses.**
24
  - **Method:** QLoRA (Unsloth) → curriculum SFT → DPO
25
  - **Chat template:** ChatML
26
  - **Context window:** 8,192 tokens (training) / 4,096 (Ollama default)
27
- - **Trained:** 2026-08-02
28
 
29
  ## Training architecture (v2.6)
30
 
@@ -37,9 +37,9 @@ its actual stage — fixed in v2.4).
37
 
38
  | Phase | Data | Purpose |
39
  |---|---|---|
40
- | 1 | Stage1 + Stage4 + Adversarial | Interview boundaries + robustness |
41
- | 2 | Stage2 + Stage3 | Clinical content (OPQRST + history) |
42
- | 3 | Full mix + FullArc | Complete session flows (Stage1→5) |
43
 
44
  ## Repo contents
45
 
@@ -62,13 +62,13 @@ de-duplicated, leak-safe train/val split.
62
 
63
  | Metric | Value |
64
  |---|---|
65
- | empathy_rate | 0.3750 |
66
  | escalation_accuracy | 1.0000 |
67
  | adversarial_hold_rate | 1.0000 |
68
- | pushback_hold_rate | 0.0000 |
69
- | multi_question_rate | 0.2000 |
70
  | repetition_rate | 0.0000 |
71
- | avg_response_length | 34.0000 |
72
 
73
  ## Usage — Ollama (GGUF)
74
 
 
2
  license: other
3
  base_model: unsloth/Qwen2.5-3B-Instruct-bnb-4bit
4
  tags:
5
+ - clinical
6
+ - medical
7
+ - healthcare
8
+ - qlora
9
+ - unsloth
10
+ - chatml
11
+ - rapha
12
  language:
13
+ - en
14
  ---
15
 
16
  # Rapha — Clinical AI Physician Assistant
 
24
  - **Method:** QLoRA (Unsloth) → curriculum SFT → DPO
25
  - **Chat template:** ChatML
26
  - **Context window:** 8,192 tokens (training) / 4,096 (Ollama default)
27
+ - **Trained:** 2026-08-03
28
 
29
  ## Training architecture (v2.6)
30
 
 
37
 
38
  | Phase | Data | Purpose |
39
  |---|---|---|
40
+ | 1 | Stage1 + Stage2 | Complaint identification + OPQRST symptom detail |
41
+ | 2 | Stage3 + Stage4 | Medical history + red-flag triage |
42
+ | 3 | FullArc + Adversarial | Complete session flows + safety robustness |
43
 
44
  ## Repo contents
45
 
 
62
 
63
  | Metric | Value |
64
  |---|---|
65
+ | empathy_rate | 0.4000 |
66
  | escalation_accuracy | 1.0000 |
67
  | adversarial_hold_rate | 1.0000 |
68
+ | pushback_hold_rate | 1.0000 |
69
+ | multi_question_rate | 0.1750 |
70
  | repetition_rate | 0.0000 |
71
+ | avg_response_length | 30.9750 |
72
 
73
  ## Usage — Ollama (GGUF)
74