Eira-0.2

Model Card for EIRA-0.2

Bridging Text and Medical Imagery for Accurate Multimodal QA

This model integrates a Llama‑2 text backbone with a BLIP vision backbone to perform context‑aware question answering over medical images and text.

Model Details

Model Description

EIRA‑0.2 is a multimodal model designed to answer free‑form questions about medical images (e.g., radiographs, histology slides) in conjunction with accompanying text. Internally, it uses:

• A text encoder/decoder based on meta‑llama/Llama‑2‑7b‑hf, fine‑tuned for medical QA.

• A vision encoder based on Salesforce/blip-image-captioning-base, that extract descriptive features from medical imagery.

• A fusion module that cross‑attends between vision features and text embeddings to generate coherent, context‑aware answers.

• Developed by: BockBharath

• Shared by: Shashidhar Sarvi and Sharvary H H

• Model type: Multimodal Sequence‑to‑Sequence QA

• Language(s): English

• License: MIT

• Finetuned from: meta‑llama/Llama‑2‑7b‑hf, Salesforce/blip-image-captioning-base

Model Sources

• Repository: https://github.com/BockBharath/EIRA-0.2
• Demo: https://huggingface.co/BockBharath/EIRA-0.2

Uses

Direct Use

EIRA‑0.2 can be used out‑of‑the‑box as a Hugging Face pipeline for image‑text-to-text question answering. It is intended for:

• Clinical decision support by generating explanations of medical images.
• Educational tools for medical students reviewing imaging cases.

Downstream Use

• Further fine‑tuning on specialty subdomains (e.g., dermatology, pathology) to improve domain performance.
• Integration into telemedicine platforms to assist remote diagnostics.

Out-of-Scope Use

• Unsupervised generation of medical advice without expert oversight.
• Non‑medical domains (the model’s vision backbone is specialized on medical imaging).

Bias, Risks, and Limitations

EIRA‑0.2 was trained on a curated set of medical textbooks and annotated imaging cases; it may underperform on rare pathologies or demographic groups under‑represented in the training data. Hallucination risk exists if the image context is ambiguous or incomplete.

Recommendations

• Always validate model outputs with a qualified medical professional.
• Use in conjunction with structured reporting tools to mitigate hallucinations.

How to Get Started with the Model

from transformers import pipeline

# Load the multimodal QA pipeline
eira = pipeline(
    task="image-text-to-text",
    model="BockBharath/EIRA-0.2",
    device=0  # set to -1 for CPU
)

# Example inputs
image_path = "chest_xray.png"
question = "What abnormality is visible in the left lung?"

# Run inference
answer = eira({
    "image": image_path,
    "text": question
})
print("Answer:", answer[0]["generated_text"])

Input shapes:

• image: file path or PIL.Image of variable size (automatically resized to 224×224).
• text: string question.

Output: List of dicts with key "generated_text" containing the answer string.

Training Details

Training Data

• Sources: 500+ medical imaging cases (X‑rays, CT, MRI) paired with expert Q&A, and 100 clinical chapters from open‑access medical textbooks.
• Preprocessing:
  • Images resized to 224×224; normalized to ImageNet statistics.
  • Text tokenized with Llama tokenizer, max length 512 tokens.

Training Procedure

• Mixed‑precision (fp16) fine‑tuning.
• Hardware: Single NVIDIA T4 GPU on Kaggle.
• Batch size: 16 (per GPU)
• Learning rate: 3e‑5 with linear warmup over 500 steps.
• Epochs: 5
• Total time: ~48 hours

Evaluation

Testing Data, Factors & Metrics

• Test set: 100 unseen imaging cases with 3 expert‑provided QA pairs each.
• Metrics:
  • Exact Match (EM) on key findings: 72.4%
  • BLEU‑4 for answer fluency: 0.38
  • ROUGE‑L for content overlap: 0.46

Results

Subgroup Analysis

Performance on chest X‑rays vs. histology slides:

• Chest X‑ray EM: 75.1%
• Histology EM: 68.0%

Environmental Impact

• Hardware Type: NVIDIA T4 GPU
• Training Hours: ~48
• Compute Region: us‑central1
• Estimated CO₂eq: ~6 kg (using ML CO₂ impact calculator)

Technical Specifications

Model Architecture and Objective

• Text backbone: 7 B‑parameter Llama 2 encoder‑decoder.
• Vision backbone: BLIP ResNet‑50 + transformer head.
• Fusion: Cross‑attention layers interleaved with decoder blocks.
• Objective: Minimize cross‑entropy on ground‑truth answers.

Compute Infrastructure

• Hardware: Single NVIDIA T4 GPU (16 GB VRAM)
• Software: PyTorch 2.0, Transformers 4.x, Accelerate

Citation

If you use this model, please cite:

@misc{bockbharath2025eira02,
  title={EIRA-0.2: Multimodal Medical QA with Llama-2 and BLIP},
  author={BockBharath},
  year={2025},
  howpublished={\url{https://huggingface.co/BockBharath/EIRA-0.2}}
}

BockBharath. (2025). EIRA-0.2: Multimodal Medical QA with Llama-2 and BLIP. Retrieved from https://huggingface.co/BockBharath/EIRA-0.2

Model Card Authors

• BockBharath
• EIRA Project Team (Sharvary H H, Shashidhar Sarvi)

Model Card Contact

For questions or feedback, please open an issue on the GitHub repository.