Model Card for 911 Operator Assistant

This model is a fine-tuned version of Mistral-7B-v0.1, designed to assist 911 operators in handling emergency calls professionally and efficiently.

Model Details

Model Description

Developed by: The model was developed using the dispatch.ipynb notebook
Model type: Fine-tuned Large Language Model
Language(s) (NLP): English
License: MIT
Finetuned from model: mistralai/Mistral-7B-v0.1

Uses

Direct Use

This model is intended to be used as an assistant for 911 operators, helping them respond to emergency calls quickly and professionally.

Out-of-Scope Use

This model should not be used as a replacement for trained 911 operators or emergency responders. It is meant to assist, not replace, human judgment in emergency situations.

Bias, Risks, and Limitations

The model may have biases based on the training data used. It should not be relied upon for making critical decisions in emergency situations without human oversight.

Recommendations

Users should always verify the model's outputs and use them in conjunction with established emergency response protocols.

How to Get Started with the Model

Use the following code to initialize the model:

from peft import PeftModel
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "mistralai/Mistral-7B-v0.1"
LORA_CHECKPOINT = "./lora_adapters/checkpoint-200/"

model, tokenizer = setup_model_and_tokenizer(BASE_MODEL)
model = PeftModel.from_pretrained(model, LORA_CHECKPOINT)
model.to(torch.device("xpu" if torch.xpu.is_available() else "cpu"))

Then, you can generate 911 operator responses by providing an input prompt:

prompt = "911 Operator: 9-1-1, what's your emergency?\nCaller: There's a fire in my kitchen!\n911 Operator:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Details

Training Data

The model was fine-tuned on a dataset of 911 call transcripts, using the "spikecodes/911-call-transcripts" dataset.

Training Procedure

Training Hyperparameters

Batch size: 4
Learning rate: 2e-5
Epochs: 7.62 (based on max_steps)
Max steps: 200
Warmup steps: 20
Weight decay: Not specified
Gradient accumulation steps: 4
Training regime: BFloat16 mixed precision

Speeds, Sizes, Times

Training time: Approximately 800.64 seconds (13.34 minutes)

Evaluation

Testing Data, Factors & Metrics

Testing Data

The model was evaluated on a validation set derived from the same dataset used for training.

Environmental Impact

Hardware Type: Intel(R) Data Center GPU Max 1100
Hours used: Approximately 0.22 hours (13.34 minutes)

Technical Specifications

Model Architecture and Objective

The model uses the Mistral-7B architecture with LoRA (Low-Rank Adaptation) for efficient fine-tuning.

Compute Infrastructure

Hardware

Intel(R) Data Center GPU Max 1100

Software

Python 3.9.18
PyTorch 2.1.0.post0+cxx11.abi
Transformers library
PEFT library
Intel Extension for PyTorch

Model Card Authors

https://github.com/spikecodes

Model Card Contact

For more information, please email me (using the contact button on my website: https://spike.codes) and refer to the repositories of the used libraries and base model.

Framework versions

PEFT 0.11.1

spikecodes
/

ai-911-operator