Retina-agent

Retina-agent is an exploratory workflow for building a retina-focused literature assistant from public biomedical articles. The repository contains notebooks and supporting files for downloading open-access PubMed Central articles, preparing training/evaluation data, testing JSONL formatting, and experimenting with local Ollama/LLaMA-Factory model workflows.

Repository Contents

NCBI_paperdownload.Rmd      # Download and process open-access PMC articles
Prepare_for_Ollama.Rmd      # Notes for building a local Ollama model
debug.Rmd                   # JSONL and dataset registration checks
Model_test/yang_model/      # Ollama Modelfile used in local experiments
report/                     # Slide/report artifacts from model experiments

What This Project Demonstrates

• Programmatic retrieval of open-access biomedical literature from NCBI/PMC.
• Dataset preparation for retina-domain language-model experiments.
• Local model experimentation with Ollama and LLaMA-Factory style data layouts.
• Practical debugging checks for ShareGPT-style JSONL training data.

Reproducibility Notes

The notebooks include paths from the original HPC and local analysis environment. Before rerunning the workflow, update those paths to match your own project directory, conda environment, NCBI API key setup, and model-output location.

Recommended local structure:

papers_retina/
  pdf/
  xml/
  metadata/
data/
  enhanced_llama3_retina_training_sharegpt.jsonl

Minimal Workflow

1. Use NCBI_paperdownload.Rmd to search PMC and download open-access retina literature.
2. Convert article text and metadata into the desired training/evaluation JSONL format.
3. Run the checks in debug.Rmd before registering the dataset with a local training framework.
4. Use Prepare_for_Ollama.Rmd and Model_test/yang_model/Modelfile for local Ollama experiments.

Notes

Only open-access articles should be downloaded and reused. Keep API keys, private datasets, and large generated model artifacts outside the repository.