Llama-3.1-8B-Chat-NVDRS-NarrativeGeneration

Model Overview

• Model Architecture: Meta-Llama-3.1
  • Input: Text (Multiple case documents: law enforcement, medical examiner, etc.)
  • Output: Text (Professional incident narrative ~600 words)
• Model Optimizations:
  • Fine-tuning: Specialized for NVDRS incident narrative generation
  • Chat Format: Llama 3.1 Instruct format compatible
  • Document Processing: Optimized for multi-document summarization
  • Production Ready: Advanced prompt engineering with detailed NVDRS formatting rules
• Intended Use Cases: Alaska Violent Death Reporting System (VDRS) for generating incident narratives to assist trained abstractors in NVDRS database submissions.
• Out-of-scope: Use outside of professional VDRS workflows. Not intended for general narrative generation.
• Release Date: June 2025
• Version: 1.0 (Alpha)
• License(s): llama3.1
• Model Developers: Alaska VDRS AI Modernization Project

Fine-tuned version of Meta-Llama-3.1-8B-Instruct specialized for generating incident narratives from multiple case documents for NVDRS reporting. Optimized for development, training, and high-precision inference with professional prompt engineering for NVDRS narrative generation.

Model Optimizations

This model was obtained by fine-tuning Meta-Llama-3.1-8B-Instruct on a curated dataset derived from NVDRS coding manual documentation and example case abstractions. The fine-tuning process enhanced the model's ability to:

• Process multiple input documents (law enforcement reports, medical examiner data, etc.)
• Generate coherent incident narratives following NVDRS guidelines
• Maintain professional tone appropriate for government reporting
• Extract and synthesize relevant information from diverse document types
• Produce ~600 word summaries suitable for CDC NVDRS database submission
• Apply privacy protection (automatic PII filtering)
• Ensure chronological event ordering in narratives
• Use professional medical terminology appropriately

Important: This model is designed to assist trained professionals and all outputs require human review by qualified Alaska VDRS abstractors.

Deployment

Use with Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "SafeAlaskans/Llama-3.1-8B-Chat-NVDRS-NarrativeGeneration"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example: Generate incident narrative from case documents
messages = [
    {
        "role": "system", 
        "content": "You are an AI assistant specialized in generating incident narratives for the Alaska Violent Death Reporting System (VDRS). Generate professional, factual narratives based on provided case documents for submission to the CDC's NVDRS database."
    },
    {
        "role": "user", 
        "content": "Based on the following case documents, generate an incident narrative for NVDRS submission:\n\nLAW ENFORCEMENT REPORT:\n{document_1_content}\n\nMEDICAL EXAMINER REPORT:\n{document_2_content}\n\nADDITIONAL DOCUMENTATION:\n{document_3_content}"
    },
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        input_ids,
        max_new_tokens=800,  # ~600 word target
        temperature=0.3,     # Lower temperature for factual content
        top_p=0.9,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id
    )

narrative = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(narrative)