A GFlowNet (Generative Flow Network) is a type of generative model trained to sample objects with probability proportional to a reward function, ensuring diversity in the generated output.

How does GFlowNet differ from Reinforcement Learning?

While standard RL aims to find a single policy that maximizes the expected reward, GFlowNet aims to find all high-reward solutions and sample them according to their quality.

What is Trajectory Balance in this context?

Trajectory Balance (TB) is a training objective for GFlowNets that ensures the flow through any given path matches the reward of the terminal state, making credit assignment more efficient.

Can this skill be used for molecule design?

Yes, one of the primary use cases for GFlowNets is generating diverse molecular graphs that satisfy specific chemical properties like synthesizability and binding affinity.

Which frameworks are supported by the GFlowNet skill?

The skill provides guidance and implementation patterns for Python using JAX, PyTorch, and the torchgfn library.

GFlowNet (Generative Flow Networks)

Name: GFlowNet (Generative Flow Networks)
Author: plurigrid

byplurigrid

•

Data Science & ML

Implements Bengio's Generative Flow Networks for diverse, reward-proportional sampling in molecule design and causal discovery.

The GFlowNet skill brings the power of Generative Flow Networks to Claude, allowing for the generation of diverse candidates in complex state spaces where traditional Reinforcement Learning or MCMC might fail. By learning to sample with probability proportional to a reward function, it enables developers and researchers to explore multiple high-reward modes rather than converging on a single global optimum. This is particularly valuable for scientific discovery, drug design, and Bayesian structure learning where finding a variety of high-quality solutions is more useful than finding just one.

Key Features

01Provides implementation patterns for Python, JAX, PyTorch, and torchgfn

027 GitHub stars

03Implements Trajectory Balance (TB) and Detailed Balance (DB) training objectives

04Enables non-iterative diverse candidate generation for complex discovery tasks

05Supports sequential decision-making for molecule synthesis and DAG sampling

06Facilitates off-policy training on diverse datasets without partition function knowledge

Use Cases

01Performing Bayesian causal discovery to identify DAG structures from observational data

02Solving combinatorial optimization problems where multiple diverse solutions are required

03Designing novel drug molecules with balanced binding affinity and drug-likeness

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi gflownet

For use in Claude.ai and ChatGPT

Download Skill