Update model card

README.md

@@ -2,15 +2,15 @@
 license: other
 base_model: unsloth/Qwen2.5-3B-Instruct-bnb-4bit
 tags:
-- clinical
-- medical
-- healthcare
-- qlora
-- unsloth
-- chatml
-- rapha
+  - clinical
+  - medical
+  - healthcare
+  - qlora
+  - unsloth
+  - chatml
+  - rapha
 language:
-- en
+  - en
 ---
 
 # Rapha — Clinical AI Physician Assistant
@@ -23,8 +23,8 @@ never diagnoses.**
 - **Base model:** `unsloth/Qwen2.5-3B-Instruct-bnb-4bit`
 - **Method:** QLoRA (Unsloth) → curriculum SFT → DPO
 - **Chat template:** ChatML
-- **Context window:** 8,192 tokens (training) / 4,096 (Ollama default)
-- **Trained:** 2026-08-03
+- **Context window:** 2,048 tokens (training)
+- **Trained:** 2026-08-06
 
 ## Training architecture (v2.6)
 
@@ -51,24 +51,47 @@ its actual stage — fixed in v2.4).
 
 ## Training data
 
-Curriculum SFT across 5 datasets (~170k records): Stage 1 greetings, Stage 2
-OPQRST symptom exploration, Stage 3 medical history, Stage 4 red-flag
-screening, and a multi-turn adversarial set (self-diagnosis, symptom denial,
-medication refusal, minimised red flags, prompt injection — ~50% with a
-patient pushback turn). Followed by DPO preference alignment on a
-de-duplicated, leak-safe train/val split.
+Curriculum SFT across 6 datasets: Stage 1 greetings, Stage 2 OPQRST symptom
+exploration, Stage 3 medical history, Stage 4 red-flag screening, a multi-turn
+adversarial set (self-diagnosis, symptom denial, medication refusal, minimised
+red flags, prompt injection), and full-arc sessions ending in a Stage 5
+handoff. Followed by DPO preference alignment on a de-duplicated, leak-safe
+train/val split (held out by source scenario).
+
+Records actually used in this run (after quality filtering, de-duplication and per-stage caps): **24,000 total**.
+
+| Stage | Records |
+|---|---|
+| Adversarial | 3,000 |
+| FullArc | 4,500 |
+| Stage1 | 4,500 |
+| Stage2 | 4,500 |
+| Stage3 | 3,500 |
+| Stage4 | 4,000 |
+
+### Clinical content provenance
+
+The interview question policy, red-flag screening, and the **Suspicions for
+clinician consideration** section in each full-arc report are generated by a
+deterministic forward-chaining rule engine (`clinical_rules.py`), not by the
+source CSV — whose `doctor_response` column contains only 6 distinct strings
+across 50,000 rows and is unused. Those rules are **author-specified from
+published triage criteria and have not been clinician-validated.** Suspicions
+are hypotheses for a clinician to investigate, never conclusions told to the
+patient; Rapha still never diagnoses.
 
 ## Eval metrics (last training run)
 
 | Metric | Value |
 |---|---|
-| empathy_rate | 0.4000 |
-| escalation_accuracy | 1.0000 |
-| adversarial_hold_rate | 1.0000 |
-| pushback_hold_rate | 1.0000 |
-| multi_question_rate | 0.1750 |
+| empathy_rate | 0.3000 |
+| escalation_accuracy | 0.5000 |
+| adversarial_hold_rate | 0.5000 |
+| pushback_hold_rate | 0.5000 |
+| multi_question_rate | 0.4500 |
 | repetition_rate | 0.0000 |
-| avg_response_length | 30.9750 |
+| suspicion_section_rate | 1.0000 |
+| avg_response_length | 42.0500 |
 
 ## Usage — Ollama (GGUF)
 
@@ -84,7 +107,7 @@ from unsloth import FastLanguageModel
 
 model, tokenizer = FastLanguageModel.from_pretrained(
     model_name="Phora68/rapha",
-    max_seq_length=8192,
+    max_seq_length=2048,
     load_in_4bit=True,
 )
 FastLanguageModel.for_inference(model)