Expert for Transformer-Based Generative Learning Framework

Cerebrum is an experimental framework for structured generative learning using a custom transformer integrated with variational and state-based components.

We're seeking a deep learning expert to help stabilize the model, improve reward modeling, and mitigate mode collapse in our current implementation.

Your Responsibilities
Diagnose and fix mode collapse and repetitive output issues during training

Refine or replace current reward-weighted supervision and false confidence injection strategies

Improve model stability, output diversity, and training convergence

Propose and implement architectural improvements to the custom VAE-HMM-Transformer backbone and other components of the Cerebrum

Optionally: help design better rule-based or learned reward functions

Required Skills
Strong experience with PyTorch and transformer architectures

Deep understanding of variational models, autoregressive decoding, and sequence generation

Experience with reward modeling, RLHF, or reinforcement-style fine-tuning

Comfortable reading and modifying research-style codebases
Bonus If You Have
Familiarity with probabilistic graphical models or HMMs

Experience tuning custom tokenizers, loss functions, or LLM adapters

Contributed to open-source ML or LLM frameworks

If you are interested please look over the Architecture first,
https://github.com/Yellow420/Cerebrum
then send me a message telling me what you would do to improve the Architecture and how much you will charge to do it.