Polysome

Polysome is a generic data generation framework designed for transforming text attributes using Large Language Models. It serves as a powerful workflow builder for chaining prompts to generate synthetic data at scale.

While originally developed for computational pathology, Polysome is domain-agnostic and focuses purely on the data generation aspect of the pipeline. It allows you to define complex, multi-step text transformation workflows using a node-based architecture.

Key Features:

• Workflow Engine: Define DAG-based pipelines to load data, process it with LLMs, and structure the output.
• Synthetic Data Generation: Ideal for creating instruction tuning datasets, rewriting reports, or extracting structured information from unstructured text.
• Prompt Chaining: distinct nodes for complex reasoning tasks, summarization, and translation.
• High Performance: Supports batch processing and Data Parallelism via vLLM.

🚀 Quick Start

Installation

Install Polysome from PyPI (minimal install, HuggingFace only):

pip install polysome

For specific engines or features:

# vLLM (Recommended for Linux + NVIDIA GPU)
pip install "polysome[vllm]"

# llama.cpp (Recommended for CPU or Apple Silicon)
pip install "polysome[llama-cpp]"

# UI / Prompt Editor
pip install "polysome[ui]"

# Install everything (for development/testing)
pip install "polysome[all]"

Convenience aliases:

• pip install "polysome[gpu]" → installs vllm stack
• pip install "polysome[cpu]" → installs llama-cpp stack

🔑 Hugging Face Authentication

Many models (like Gemma-3) are gated and require authentication. Before running workflows with these models, log in via the CLI:

pip install huggingface_hub
huggingface-cli login

Alternatively, set the HF_TOKEN environment variable.

Create Your First Project

Initialize a new project with example workflows and prompts:

polysome init my-analysis
cd my-analysis

This creates:

• workflows/ - Example workflow configurations
• prompts/ - Prompt templates for your tasks
• data/ - Directory for input data (with example file)
• output/ - Results will be saved here

Run a Workflow

# Run the example workflow
polysome run workflows/basic_text_generation.json

# Run with custom settings
polysome run workflows/my_workflow.json --log-level DEBUG

Customize for Your Use Case

1. Edit your workflow (workflows/basic_text_generation.json):

   • Change the model name (default: google/gemma-3-4b-it)
   • Adjust generation parameters
   • Add or remove processing nodes

2. Customize prompts (prompts/simple_qa/):

   • system_prompt.txt: Define the AI's role
   • user_prompt.txt: Template with variables like {{ question }}
   • few_shot.jsonl: Example inputs and outputs

3. Prepare your data (data/input.json):

   • Format as JSON with fields matching your prompt variables

4. Run and iterate:

   polysome run workflows/basic_text_generation.json

Using Docker (Alternative)

For reproducible environments or deployment:

# Run with Docker
docker run --rm --gpus all \
  -v ./data:/data \
  -v ./output:/output \
  -v ./workflows:/workflows \
  -v ./prompts:/prompts \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e WORKFLOW_PATH=/workflows/basic_text_generation.json \
  ghcr.io/computationalpathologygroup/polysome:latest

Note: We recommend mounting your Hugging Face cache to avoid re-downloading models.

For detailed Docker usage, see docs/docker_container.md.

🐍 Programmatic API

For advanced users, you can also use Polysome programmatically:

from polysome.workflow import Workflow

# Load and run a workflow
workflow = Workflow("workflows/my_workflow.json")
success = workflow.run()

# Access results
print(f"Results saved to: {workflow.output_dir}")

🛠️ Contributing & Development

Want to contribute or modify Polysome? See CONTRIBUTING.md for guidelines.

Local Development Setup

git clone https://github.com/computationalpathologygroup/Polysome.git
cd Polysome

# Install in development mode with dependencies
pip install -e ".[gpu-dev]"  # For GPU development
# OR
pip install -e ".[cpu-dev]"  # For CPU development

Running Tests

pytest
pytest --cov=polysome  # With coverage

🧩 Workflow Configuration

Workflows are defined in JSON files (DAGs) located in the workflows/ directory. They control how data is loaded, processed by LLMs, and saved.

For a detailed guide on creating nodes and configuring JSONs, see docs/text_preprocessing.md.

🎨 Prompt Engineering

Polysome includes a Streamlit-based Prompt Editor to help you design, manage, and test Jinja2 templates for your LLM tasks.

# Run the editor
polysome-gui

For a user guide on managing templates and few-shot examples, see docs/prompt_editor.md.

⚡ High Performance Inference

For large-scale processing, Polysome supports Data Parallelism using vLLM to distribute batches across multiple GPUs.

To enable this, use the vllm_dp engine in your workflow configuration. See docs/data_parallelism.md for setup instructions and performance tuning.

📚 Documentation Index

• Text Preprocessing & Workflows
• Docker Container Guide
• Prompt Editor Guide
• Data Parallelism Guide

📄 Citation

This framework was originally developed to support visual instruction tuning. If you use this code to generate data for such models, please consider citing the following paper:

@inproceedings{moonemans2025open,
  title={Democratizing Pathology Co-Pilots: An Open Pipeline and Dataset for Whole-Slide Vision-Language Modeling},
  author={Sander Moonemans and Sebastiaan Ram and Fr{\'e}d{\'e}rique Meeuwsen and Carlijn Lems and Jeroen van der Laak and Geert Litjens and Francesco Ciompi},
  booktitle={Submitted to Medical Imaging with Deep Learning},
  year={2025},
  url={https://openreview.net/forum?id=aGPowreqPi},
  note={under review}
}