Skip to content
/ rbio Public

Public repo for rbio, a biologically-informed reasoning model trained on virtual cell models as verifiers

License

Notifications You must be signed in to change notification settings

czi-ai/rbio

Repository files navigation

rbio-1

rbio Model Architecture

rbio-1 is a reasoning model that was trained using predictions from Virtual Cell Models (VCMs) as soft verification during training.

The Rbio model is based on Qwen2.5-3B-Instruct, which is licensed under the Qwen Research License. All modifications and further developments by CZI are released under the MIT License.

Citation

Ana-Maria Istrate, Fausto Milletari, Fabrizio Castrotorres, Jakub Tomczak, Michaela Torkar, Donghui Li, Theofanis Karaletsos. rbio-1 - training scientific reasoning LLMs with biological world models as soft verifiers (2025) bioRxiv. DOI: https://doi.org/10.1101/2025.08.18.670981

Model Variants

Rbio includes several variants based on the type of data or model used as a verifier during reinforcement learning.

Model Variant Name Task, Purpose, or Description Access URL or AWS Download link
Rbio1-EXP Post-trained using direct experimental data as a "hard verifier” for maximum accuracy on related tasks. s3://czi-rbio/rbio1-EXP/
Rbio1-MLP Post-trained using a task-specific MLP as a "soft verifier”, demonstrating knowledge transfer from a smaller world model. s3://czi-rbio/rbio1-MLP/
Rbio1-TF Post-trained using signals (e.g., PMI scores) from the Transcriptformer foundation model as a “soft verifier”. s3://czi-rbio/rbio1-TF/
Rbio1-GO Post-trained using the Gene Ontology (GO) knowledge base as a “soft verifier” guiding the model with established biological facts via ROUGE metric. Includes information from all subsets of GO Ontology: GO-F (Molecular Function), GO-P (Biological Processes) and GO-C (Cellular Component) s3://czi-rbio/rbio1-GO/
Rbio1-GO-C Post-trained using the Gene Ontology (GO) knowledge base Cellular Component (GO-C) as a “soft verifier” guiding the model with established biological facts via ROUGE metric. s3://czi-rbio/rbio1-GO-C/
Rbio1-GO-F Post-trained using the Gene Ontology (GO) knowledge base Mollecular Function (GO-F) as a “soft verifier” guiding the model with established biological facts via Rouge metric. s3://czi-rbio/rbio1-GO-F/
Rbio1-GO+EXP Post-trained using both experimental data acting as a “hard verifier” on the task at hand and Gene Ontology (GO-all: GO-P + GO-C + GO-F) knowledge base as a “soft verifier” for biological facts consistency. s3://czi-rbio/rbio1-GO+EXP/
Rbio1-TF+EXP Post-trained using both experimental data acting as a “hard verifier” on the task at hand and Transcriptformer foundation model as “soft verifier” using PMI scores. s3://czi-rbio/rbio1-TF+EXP/
Rbio1-TF+GO+EXP Post-trained using: experimental data acting as a “hard verifier” on the task at hand; Transcriptformer foundation model as “soft verifier” using PMI scores; and Gene Ontology (GO-all: GO-P + GO-C + GO-F) knowledge base as a “soft verifier” for biological facts consistency. s3://czi-rbio/rbio1-TF+GO+EXP/
Rbio1-TF+GO+MLP Post-trained using: an MLP acting as a "soft verifier” of world-knowledge as seen through the lens of a smaller model; Transcriptformer foundation model as “soft verifier” using PMI scores; and Gene Ontology (GO-all: GO-P + GO-C + GO-F) knowledge base as a “soft verifier” for biological facts consistency/ s3://czi-rbio/rbio1-TF+GO+MLP/
Rbio1-TF+GO+MLP+EXP Post-trained using: experimental data acting as a “hard verifier” on the task at hand; Transcriptformer foundation model as “soft verifier” using PMI scores; Gene Ontology (GO-all: GO-P + GO-C + GO-F) knowledge base as a “soft verifier” for biological facts consistency; and MLP as a "soft verifier” of world-knowledge as rendered via a smaller model. s3://czi-rbio/rbio1-TF+GO+MLP+EXP/

Usage

We recommend creating a virtual env with:

python3 -m venv rbio-env
source rbio-env/bin/activate
pip3 install -r requirements.txt

1. Inference Scripts

The inference scripts will run an rbio model version on a list of user-provided questions. The script will automatically download the model weights from AWS S3.

The model arguments are:

argument description default_value
base_model_name base model name Qwen/Qwen2.5-3B-Instruct
rbio_model_ckpt rbio_model_variation rbio_TF_ckpt
results_output_folder optional folder where to save the results predictions
results_output_filename optional filename for the results results.csv

2. Training Scripts

The training scripts demonstrate a minimal example of RBIO training with soft verification. This implementation uses soft verification against a simplified biological perturbation model based on a multi-layer perceptron (MLP). For detailed instructions on running the training pipeline, see the README.md file in the training/ subdirectory.

Code of Conduct

This project adheres to the Contributor Covenant code of conduct. By participating, you are expected to uphold this code. Please report unacceptable behavior to [email protected].

Reporting Security Issues

If you believe you have found a security issue, please responsibly disclose by contacting us at [email protected].

About

Public repo for rbio, a biologically-informed reasoning model trained on virtual cell models as verifiers

Resources

License

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 5

Languages