I poured beer for a living. ML obsessed neural network neurologist.
| Machine Learning | 73 |
|---|---|
| Computer Vision | 0 |
| Data Science | 55 |
| NLP & LLMs | 77 |
| AI Agents | 0 |
| MLOps | 59 |
Each node is one domain, scored only from evidence in these public repositories. Select one to see why.
Deep, repeated, well-engineered work in this domain.
30 public repositories → 22 with AI work, evaluated. The other 8 (repositories without AI code, forks and empty repositories) are not scored and have no effect on the result. How repositories are selected
Ordered by how much each project adds to the Atlas score, not by its own score. A project with evidence in several domains can rank above one with a slightly higher project score. Choose a domain to focus on the projects with meaningful evidence there.
21% of Atlas Score
The repository provides a library for training JumpReLU sparse autoencoders (SAEs) per transformer layer with an event-aware optimizer scheduler and augmented Lagrangian sparsity control. It includes tools for activation capture, checkpointing, and diagnostics, demonstrated by training a 35-layer SAE atlas for Gemma-4-E2B-it.
14% of Atlas Score
Sub-Zero is an end-to-end interpretability and steering system for LLMs that identifies and attenuates corporate/refusal directions in model weights through SVD-based probing and gradient-masked fine-tuning, accompanied by a pipeline for building a detailed Brain Atlas.
12% of Atlas Score
AECS is a PyTorch learning rate scheduler that adjusts the learning rate in real time based on training signals such as gradient norm and loss, switching between four modes (BASELINE, RECOVERY, EXPLORE, STABILIZE) using event-driven logic.
9% of Atlas Score
The repository provides a Dockerfile and pinned dependency files to build a reproducible base image for running the Qwen/Qwen3.8-27B model, eliminating environment setup by guaranteeing CUDA/PyTorch compatibility and pre‑installing required libraries.
8% of Atlas Score
bellapy is a Python library and CLI toolkit designed for data preparation in machine learning workflows. It provides functions for deduplication, cleaning, format conversion, and other common data processing tasks.
7% of Atlas Score
This repository implements a gather-matmul-scatter overlay for the `deep_chaos_scheduler` to optimize training by computing only on active channels. It includes the overlay implementation, tests for correctness, and benchmarks to measure performance impact.
