Bio-Medical-Llama-3-2-1B-CoT-012025

This model is a fine-tuned version of Llama-3.2-1B-Instruct on our custom "BioMedData" dataset, enhanced with 625,000 examples, including 25,000 chain-of-thought (CoT) instruction samples to strengthen reasoning capabilities. It is specifically optimized for the Healthcare & Lifesciences (HLS) domain.

Model details

Model Name: Bio-Medical-Llama-3-2-1B-CoT-012025

Base Model: Llama-3.2-1B-Instruct

Parameter Count: 1 billion

Training Data: Custom high-quality biomedical dataset with 625,000 examples, including 25,000 CoT instructions.

Number of Entries in Dataset: 625,000

Dataset Composition: The dataset comprises a mix of synthetic, manually curated, and reasoning-focused entries, ensuring comprehensive coverage of biomedical knowledge and logical reasoning.

Model description

The Bio-Medical-Llama-3-2-1B-CoT-012025 model is a lightweight yet powerful language model tailored for:

• Generating domain-specific content in healthcare and biomedical fields.
• Answering complex questions requiring step-by-step reasoning using CoT.
• Supporting researchers, clinicians, and students in their respective biomedical endeavors.

This model is fine-tuned to provide interpretability and improved logical coherence through its enhanced CoT capabilities.

Evaluation Metrics

Bio-Medical-Llama-3-2-1B-CoT-012025 has been evaluated using the Eleuther AI Language Model Evaluation Harness framework on tasks including:

• medmcqa
• medqa_4options
• mmlu_anatomy
• mmlu_clinical_knowledge
• mmlu_college_biology
• mmlu_college_medicine
• mmlu_medical_genetics
• mmlu_professional_medicine
• pubmedqa

Results show consistent performance improvements over general-purpose models of similiar size, particularly in tasks requiring reasoning.

Intended uses & limitations

Intended Uses:

1. Research Support: Assisting researchers with reasoning and data extraction from biomedical texts.
2. Clinical Decision Support: Offering logical and evidence-based information to aid decision-making.
3. Educational Tool: Serving as a learning resource for understanding complex biomedical concepts.

Limitations and Ethical Considerations:

• Biases: The model may reflect biases from the training data, despite efforts to mitigate them.
• Accuracy: Responses should be cross-verified with reliable sources in critical scenarios.
• Ethical Use: The model should augment professional expertise and not replace it, especially in high-stakes applications.

How to use

import transformers
import torch

model_id = "ContactDoctor/Bio-Medical-Llama-3-2-1B-CoT-012025"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are an expert trained on healthcare and biomedical domain!"},
    {"role": "user", "content": "What are the differential diagnoses for a patient presenting with shortness of breath and chest pain?"},
]

prompt = pipeline.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    prompt,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])

License

This model is licensed under the Bio-Medical-Llama-3-2-1B-CoT-012025 (Non-Commercial Use Only). Please review the terms and conditions before using the model.

Contact Information

For further information, inquiries, or issues related to Bio-Medical-Llama-3-2-1B-CoT-012025, please contact:

Email: [email protected]

Website: https://www.contactdoctor.in

Training hyperparameters

The following hyperparameters were used during training:

• Learning Rate: 0.0002
• Train Batch Size: 8
• Eval Batch Size: 4
• Seed: 42
• Gradient Accumulation Steps: 8
• Total Train Batch Size: 32
• Optimizer: Adam with betas=(0.9, 0.999) and epsilon=1e-08
• LR Scheduler Type: Cosine
• LR Scheduler Warmup Ratio: 0.03
• Training Steps: 2000
• Mixed Precision Training: Native AMP

Framework versions

• PEFT: 0.11.0
• Transformers: 4.40.2
• Pytorch: 2.1.2
• Datasets: 2.19.1
• Tokenizers: 0.19.1

Citation

If you use Bio-Medical-Llama-3-2-1B-CoT-012025 in your research or applications, please cite it as follows:

@misc{ContactDoctor_Bio-Medical-Llama-3.2-1B-CoT-012025,
  author = {ContactDoctor},
  title = {Bio-Medical-Llama-3-2-1B-CoT-012025: A Reasoning-Enhanced Biomedical Language Model},
  year = {2025},
  howpublished = {https://huggingface.co/ContactDoctor/Bio-Medical-Llama-3-2-1B-CoT-012025},
}