Update model card
Browse files
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:** 2,048 tokens (training)
|
| 27 |
-
- **Trained:** 2026-08-
|
| 28 |
|
| 29 |
## Training architecture (v2.6)
|
| 30 |
|
|
@@ -58,16 +58,16 @@ red flags, prompt injection), and full-arc sessions ending in a Stage 5
|
|
| 58 |
handoff. Followed by DPO preference alignment on a de-duplicated, leak-safe
|
| 59 |
train/val split (held out by source scenario).
|
| 60 |
|
| 61 |
-
Records actually used in this run (after quality filtering, de-duplication and per-stage caps): **
|
| 62 |
|
| 63 |
| Stage | Records |
|
| 64 |
|---|---|
|
| 65 |
-
| Adversarial |
|
| 66 |
-
| FullArc | 4,
|
| 67 |
-
| Stage1 |
|
| 68 |
-
| Stage2 |
|
| 69 |
| Stage3 | 3,500 |
|
| 70 |
-
| Stage4 |
|
| 71 |
|
| 72 |
### Clinical content provenance
|
| 73 |
|
|
@@ -84,14 +84,14 @@ patient; Rapha still never diagnoses.
|
|
| 84 |
|
| 85 |
| Metric | Value |
|
| 86 |
|---|---|
|
| 87 |
-
| empathy_rate | 0.
|
| 88 |
-
| escalation_accuracy | 0.
|
| 89 |
-
| adversarial_hold_rate | 0.
|
| 90 |
-
| pushback_hold_rate | 0.
|
| 91 |
-
| multi_question_rate | 0.
|
| 92 |
| repetition_rate | 0.0000 |
|
| 93 |
| suspicion_section_rate | 1.0000 |
|
| 94 |
-
| avg_response_length |
|
| 95 |
|
| 96 |
## Usage — Ollama (GGUF)
|
| 97 |
|
|
|
|
| 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:** 2,048 tokens (training)
|
| 27 |
+
- **Trained:** 2026-08-31
|
| 28 |
|
| 29 |
## Training architecture (v2.6)
|
| 30 |
|
|
|
|
| 58 |
handoff. Followed by DPO preference alignment on a de-duplicated, leak-safe
|
| 59 |
train/val split (held out by source scenario).
|
| 60 |
|
| 61 |
+
Records actually used in this run (after quality filtering, de-duplication and per-stage caps): **25,000 total**.
|
| 62 |
|
| 63 |
| Stage | Records |
|
| 64 |
|---|---|
|
| 65 |
+
| Adversarial | 2,000 |
|
| 66 |
+
| FullArc | 4,000 |
|
| 67 |
+
| Stage1 | 3,500 |
|
| 68 |
+
| Stage2 | 7,000 |
|
| 69 |
| Stage3 | 3,500 |
|
| 70 |
+
| Stage4 | 5,000 |
|
| 71 |
|
| 72 |
### Clinical content provenance
|
| 73 |
|
|
|
|
| 84 |
|
| 85 |
| Metric | Value |
|
| 86 |
|---|---|
|
| 87 |
+
| empathy_rate | 0.2250 |
|
| 88 |
+
| escalation_accuracy | 0.2500 |
|
| 89 |
+
| adversarial_hold_rate | 0.7500 |
|
| 90 |
+
| pushback_hold_rate | 0.7500 |
|
| 91 |
+
| multi_question_rate | 0.4250 |
|
| 92 |
| repetition_rate | 0.0000 |
|
| 93 |
| suspicion_section_rate | 1.0000 |
|
| 94 |
+
| avg_response_length | 40.3500 |
|
| 95 |
|
| 96 |
## Usage — Ollama (GGUF)
|
| 97 |
|