Model Card for MamaBot-Llama

MamaBot-Llama is an opensource fine-tuned large language model developed by HelpMum to assist with maternal healthcare by providing accurate and reliable answers to questions about pregnancy, maternal and child health. The model has been fine-tuned on Llama 3.1 8b using a dataset of maternal healthcare questions and answers.

Model Details

Developed by: HelpMum
Shared by : HelpMum
Model type: Causal Language Model (Llama 3.1 8b)
Language(s) (NLP): English
License: Apache-2.0
Finetuned from model: Llama 3.1 8b

Model Sources

Repository: MamaBot-Llama on Hugging Face

Uses

Direct Use

MamaBot-Llama can be directly used to provide answers to maternal healthcare questions, offering guidance and support to mothers during pregnancy, maternal and child health.

Downstream Use

The model can be integrated into healthcare applications, chatbots, or other systems that aim to provide maternal healthcare support.

Out-of-Scope Use

The model is not intended for use in medical diagnosis or treatment without the supervision of a qualified healthcare professional. It should not be used for malicious purposes or misinformation.

Bias, Risks, and Limitations

The model was trained on a specific dataset related to maternal healthcare. While it aims to provide accurate and supportive information, users should be aware of the following:

Bias: The model may reflect biases present in the training data, which could affect the quality and impartiality of the responses.
Risks: Users should not rely solely on the model for critical medical decisions. Always consult with a healthcare professional for medical advice.
Limitations: The model's responses are based on the data it was trained on and may not cover all possible scenarios or latest medical guidelines.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. It is recommended to use the model as a supplementary tool and not as a primary source of medical advice.

How to Get Started with the Model

Use the code below to get started with the model (make sure you have access to the model).

!pip install -q -U transformers bitsandbytes

from huggingface_hub import HfFolder
HfFolder.save_token('hf_...')  

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('HelpMumHQ/MamaBot-Llama')
model = AutoModelForCausalLM.from_pretrained('HelpMumHQ/MamaBot-Llama')

def generate_response(user_message):
    tokenizer.chat_template = "{%- for message in messages %}{{ bos_token + '[INST] ' + message['content'] + ' [/INST]' if message['role'] == 'user' else ' ' + message['content'] + ' ' + eos_token }}{%- endfor %}"
    messages = [{"role": "user", "content": user_message}]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors='pt', truncation=True).to("cuda")
    outputs = model.generate(**inputs, max_length=150, num_return_sequences=1)
    text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = (text[text.find('[/INST]') + len('[/INST]'):text.find('[INST]', text.find('[/INST]') + len('[/INST]'))] if text.find('[INST]', text.find('[/INST]') + len('[/INST]')) != -1 else text[text.find('[/INST]') + len('[/INST]'):]).strip().split('[/INST]')[0].strip()
    return response

# Sample usage
user_message = "Why might mothers not realize they are already pregnant in the first two weeks?"
response = generate_response(user_message)
print(response)

Training Details

Training Data

The training data consists of a HelpMum-created dataset of maternal healthcare questions and answers covering all stages of pregnancy up to birth.

Training Procedure

Preprocessing

The dataset was cleaned and formatted to align with the required input format for the model.

Training Hyperparameters

Training regime: torch.bfloat16
Optimizer: paged_adamw_32bit
Learning rate: 2e-4

Evaluation

Testing Data, Factors & Metrics

Testing Data

The testing data is a subset of the training dataset, split into training and testing sets.

Factors

The evaluation considered the training and validation losses.

Metrics

The model was evaluated based on training loss and validation loss metrics.

Results

Training Loss: 0.4654
Validation Loss: 0.5168

Summary

The model showed consistent performance with a training loss of 0.4654 and a validation loss of 0.5168, indicating its effectiveness in answering maternal healthcare questions.

Environmental Impact

Hardware Type: GPU

Technical Specifications

Model Architecture and Objective

The model is based on the Llama 3.1 8b architecture and aims to provide accurate and supportive responses to maternal healthcare questions.

Compute Infrastructure

Hardware

The model was trained using GPUs to handle the computational load of fine-tuning a large language model.

Software

The training and inference were conducted using the Hugging Face Transformers library and other associated tools.

Citation

BibTeX:

@misc{mamabot-llama,
  author = {HelpMum},
  title = {MamaBot-Llama},
  year = {2024},
  howpublished = {\url{https://huggingface.co/HelpMumHQ/MamaBot-Llama}},
}

APA:

HelpMum. (2024). MamaBot-Llama. Retrieved from https://huggingface.co/HelpMumHQ/MamaBot-Llama

Model Card Contact

For more information, please contact tech@helpmum.org.