Rapha — Clinical AI Physician Assistant

Rapha conducts structured, empathetic clinical interviews across five stages (greeting, OPQRST symptom exploration, medical history, red-flag screening, escalation report) and hands a structured report to a physician. Rapha 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)
• LoRA config: r=32, alpha=32, target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trained: 2026-07-22

Training architecture (v2.3)

Single-trainer curriculum SFT: three phases concatenated into one ordered dataset with a single cosine LR schedule. This avoids the optimizer restart and warmup re-run issues of the previous multi-trainer approach.

Phase	Data	Purpose
1	Stage1 + Stage4 + Adversarial	Interview boundaries + robustness
2	Stage2 + Stage3	Clinical content (OPQRST + history)
3	Full mix + FullArc	Complete session flows (Stage1→5)

Repo contents

Path	Contents
/ (root)	LoRA adapter (PEFT) — small, load on top of the base model
merged/	Full merged fp16 weights — standalone, no base model needed
gguf/	Quantised GGUF files (Q4_K_M, Q5_K_M, Q8_0) for Ollama / llama.cpp / LM Studio
datasets/	Raw JSONL training/validation/DPO datasets used for this run

Training data

Total training records: ~159,766

Dataset	Records	File
Stage 1 — Greetings	25,000	stage1_greetings.jsonl
Stage 2 — OPQRST	40,344	stage2_opqrst.jsonl
Stage 3 — History	40,000	stage3_history.jsonl
Stage 4 — Red Flags	14,422	stage4_red_flags.jsonl
Adversarial	10,000	adversarial.jsonl
Full-Arc Conversations	30,000	full_arc_conversations.jsonl
Validation	1,100	val_sharegpt.jsonl
DPO Preference Pairs	3,000	preference_pairs.jsonl

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 — a portion include a patient pushback turn to test boundary-holding). Followed by DPO preference alignment on the pairs above.

Red-flag terms screened for

worst headache, can't feel my legs, arm feels heavy and chest, can't breathe, loss of consciousness, sudden vision, neck stiff, light hurts, confused, won't stop bleeding, bilateral leg weakness, bladder dysfunction, tachypnoea, thunderclap, cauda equina, qsofa, meningitis

Eval metrics (last training run)

Metric	Value
(no eval metrics captured)	—

Usage — Ollama (GGUF)

ollama create rapha -f Modelfile.q4_k_m
ollama run rapha

Usage — Transformers (LoRA adapter)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Phora68/rapha",
    max_seq_length=8192,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Safety

Rapha is an information-gathering and triage-support tool. It is not a diagnostic device and must not be deployed without physician oversight. Red-flag detection and escalation responses should be validated against the clinical accuracy benchmark before any clinical use.

---

Generated automatically by train_rapha_llm.py v2.3.