---

language:

• bn
• en

license: apache-2.0

tags:

• medical
• healthcare
• clinical-triage
• diagnostic-support
• bengali
• banglish
• gguf
• llama-cpp
• ollama
• lm-studio
• rag

datasets:

• Irtisum/bengali-medical-triage-conversations

pipeline_tag: text-generation

---

🩺 Aurora V3.2 (GGUF): Multilingual Clinical Triage & Screening Assistant (8B)

Aurora-V3.2-GGUF is an 8B parameter quantized instruction-tuned clinical screening model designed for multilingual healthcare triage in Bengali (বাংলা), Banglish (Phonetic Romanized Bengali), and English.

It is trained on the Irtisum/bengali-medical-triage-conversations dataset, incorporating 750+ contrastive hard-negative pairs to differentiate overlapping acute tropical fevers (Dengue, Malaria, Typhoid, Chikungunya, Hepatitis E) without diagnostic overcalling.

---

📊 Benchmark & Evaluation Results

Aurora V3.2 was evaluated across acute disease scenarios in Bengali, Banglish, and English. The results demonstrate the power of combining Instruction Fine-Tuning (SFT) with Clinical RAG Grounding:

| Configuration | Script / Language Routing | Diagnostic Differential Accuracy |

|:---|:---:|:---:|

| Baseline LLM (Zero-Shot) | 35.0% (Broken Banglish) | < 30.0% (Severe Overcalling Bias) |

| Aurora V3.2 (SFT Parametric) | 90.0% | 50.0% |

| Aurora V3.2 + Clinical RAG Grounding | 92.5% | 95.0% 🚀 (+45.0% Accuracy Boost) |

🔬 Key Clinical Insights

• SFT Role: Fine-tuning establishes robust conversational triage behavior, clinical empathy, strict script routing (Banglish $\rightarrow$ Bengali script), and red-flag danger sign escalation.
• RAG Grounding Role: Connecting the model to structured clinical disease cards (WHO/CDC criteria) eliminates diagnostic drift, raising differential accuracy to 95.0%.

---

🌟 Clinical Triage Behavior

1. Active History Taking: Asks one targeted discriminatory question per turn (2–4 turns total) before providing an assessment.
2. Bilingual Script Routing:

- Bengali Query $\rightarrow$ Replies in Bengali script.

- Banglish Query ("amar 3 din dhore jor ar matha betha") $\rightarrow$ Replies in Bengali script.

- English Query $\rightarrow$ Replies in English.

1. Emergency Escalation: Automatically surfaces danger signs (severe bleeding, circulatory shock, respiratory distress, severe dehydration) before the differential ranking.
2. Zero-Prescription Safety: Strictly avoids prescribing medications, dosages, or unverified home remedies. Always advises consulting a licensed healthcare professional.

---

🚀 How to Run Locally

1. LM Studio (Recommended)

1. Search for Irtisum/Aurora-V3.2-GGUF inside LM Studio.
2. Download Aurora-V3.2-Q4_K_M.gguf.
3. Load the model and chat directly!

2. Ollama

Create a Modelfile:

FROM ./Aurora-V3.2-Q4_K_M.gguf
PARAMETER temperature 0.2
PARAMETER top_p 0.9
SYSTEM """You are an empathetic, clinical AI triage assistant for Bengali, Banglish, and English. Always ask one relevant follow-up question per turn. Never prescribe medicine."""

Then run:

ollama create aurora-v3.2 -f Modelfile
ollama run aurora-v3.2

3. llama.cpp CLI

./llama-cli -m Aurora-V3.2-Q4_K_M.gguf \
  -p "User: amar 3 din dhore jor ar matha betha korche\nAssistant:" \
  -n 256 --temp 0.2

---

📚 Training Dataset

This model was trained on the open-source dataset:

👉 Irtisum/bengali-medical-triage-conversations

---

⚖️ Clinical Safety Disclaimer

> DISCLAIMER: Aurora V3.2 is an experimental AI research model for academic and clinical triage benchmarking. It is not a certified medical device and must not be used as a substitute for professional medical diagnosis or clinical decision-making.