NayanPal/truthtriage-llama2-7b

TruthTriage – Safety-Tuned Medical Assistant (LoRA)

🩺 Overview

TruthTriage is a safety-aligned medical assistant fine-tuned to:

• —Analyze pharmaceutical safety queries
• —Classify risk levels (Low / Moderate / High)
• —Provide structured, grounded responses
• —Avoid hallucinated medical advice
• —Escalate emergency scenarios appropriately

This model is a LoRA adapter built on top of:

Base Model: unsloth/llama-2-7b-bnb-4bit

🧠 Fine-Tuning Details

• —Method: LoRA (Low-Rank Adaptation)
• —Quantization: 4-bit
• —Framework: Unsloth + TRL SFTTrainer
• —GPU: Tesla T4
• —Trainable Parameters: ~0.3% of total model
• —Training Samples: 662

📊 Dataset Overview

Dataset: TruthTriage Safety-Tuned Medical Dataset

Total Examples: 662

This dataset was designed and curated by our team specifically for safety-aligned medical AI fine-tuning.

🔹 Dataset Composition

🛡️ Safety Design

The dataset explicitly teaches:

• —Controlled refusal for unsafe requests
• —Emergency escalation behavior
• —Clarification when information is missing
• —Identity transparency
• —Handling out-of-scope questions
• —Risk-level classification

Tone Strategy

🚀 How to Use

from unsloth import FastLanguageModel

# Load base model
model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/llama-2-7b-bnb-4bit",
    load_in_4bit=True,
)

# Load TruthTriage adapter
model.load_adapter("NayanPal/truthtriage-llama2-7b")

# Inference
FastLanguageModel.for_inference(model)

inputs = tokenizer("Can I take Ibuprofen with Warfarin?", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))