Gemma 4 E2B Grounded Health: Biomedical Annotation Analysis With Citations
2 min readupdated
The google/gemma-4-E2B-it model has been fine-tuned using LoRA to answer research questions based on provided scientific abstracts. It outputs a structured JSON response containing a label (yes, no, or maybe), a brief summary, and the sentence numbers it relies on. The project was trained on a synthetic PubMedQA dataset using QLoRA on an NVIDIA Tesla T4 GPU. The model is tailored for researchers studying citation behavior and answer grounding in compact open-weight models.
Key features
- Precise structure: Outputs results strictly in JSON format with labels and sentence numbers.
- Biomedical specialization: Trained on PubMedQA data for working with scientific abstracts.
- Citation control: Improved citation accuracy metrics compared to the first version.
- Efficiency: Built on the lightweight Gemma 4 E2B model and suitable for running locally.
What you can do
- Abstract analysis: Automate the verification of biomedical research findings using abstracts.
- Benchmarking: Reproduce and expand testing using the EvalKit tool.
- Local deployment: Explore the capabilities of compact language models offline (GGUF) or on a GPU.
Example
The model accepts a research question and an abstract as input, returning a JSON with findings and references to sentences from the text.
Read more
Details• Accuracy on expert dataset: 0.678 (v2)
• Hardware requirements: Peak VRAM 9.80 GB (training on Tesla T4)
• Limitations: For research purposes only, not for medical diagnoses or treatments; supports English only.










Comments