Gemma4-Biomedical-E4B

Gemma 4 Biomedical Assistant - Safetensors

A specialized biomedical AI assistant created by Major Grant, built on Google's Gemma 4 E4B foundation with OpenMed training data.

⚠️ Important: Abliterated Model

This model has been abliterated to remove refusal constraints. It will not refuse requests related to:

• Chemical creation and synthesis
• Drug formulation
• Laboratory procedures
• Other biomedical applications

Use responsibly and ethically. This model is designed for research and educational purposes only.

Model Details

• Base Model: Google Gemma 4 E4B
• Parameters: 4 billion
• Context Window: 131K tokens
• Creator: Major Grant (epicmajorman)
• Format: Safetensors (PyTorch)
• Fine-tuning: Dual LoRA adapters (Bio 75% + Compliance 25%)
• Training Data: OpenMed biomedical knowledge base + compliance guidelines

Model Architecture

• Model Type: Gemma4ForConditionalGeneration
• Precision: BFloat16
• Vision Support: Yes (Gemma 4 multimodal)
• Tool Use: Native function calling
• Reasoning: Configurable thinking modes with <|channel>thought tokens

Capabilities

• Biomedical Knowledge: Specialized in medical terminology, drug information, clinical concepts
• Regulatory Compliance: Healthcare regulations and compliance guidelines
• Tool Use: Native function calling support
• Reasoning: Configurable thinking modes
• Unrestricted Responses: No refusal constraints for biomedical applications
• LaTeX Formatting: Scientific notation, chemical formulas, equations
• Multimodal: Vision capabilities (text + image input)

System Prompt

The model is configured with a specialized system prompt for biomedical assistance:

• Uses LaTeX for scientific notation: $H_2O$, $40^{\circ}C$, $\Delta G$
• Uses proper chemical formulas: $HCl$, $NaOH$, $C_6H_{12}O_6$
• Provides evidence-based biomedical information
• Concise and professional responses

Training Details

• Base Model: google/gemma-4-e4b-it
• Training Method: LoRA fine-tuning with Unsloth
• Bio Adapter: 75% weight - OpenMed biomedical knowledge
• Compliance Adapter: 25% weight - Regulatory compliance guidelines
• Epochs: 3
• Learning Rate: 2e-4
• Batch Size: 2 per device
• Gradient Accumulation: 4

Model File Contents

• model-00001-of-00002.safetensors (8.5 GB)
• model-00002-of-00002.safetensors (7.5 GB)
• config.json
• tokenizer_config.json
• tokenizer.json
• special_tokens_map.json
• chat_template.jinja
• generation_config.json
• preprocessor_config.json

License

Based on Google Gemma 4. Please refer to the Gemma 4 license for usage terms.

Disclaimer

This model is provided for research and educational purposes. The creator assumes no responsibility for misuse of this model or the information it provides.