Exploring Next-Concept Modeling (NCM) & Transformer Interpretability: Repository of 20 Exploratory Studies & Controls [P]

Hi everyone, I wanted to share an open-source repo collecting 20 exploratory studies focused on Next-Concept Modeling (NCM), mechanistic interpretability (like IOI in GPT-2), and model controls.

Link: https://github.com/Quantix-eng/ncm-experiment-notes

Project Focus

This is an independent project documenting empirical experiments on how transformers process and edit internal features. No performance breakthroughs here - just honest documentation of observations alongside rigorous controls.

The repo covers:

• NCM on structured tasks (like Sudoku)
• Mechanistic interpretability in GPT-2
• Toy parallel models

Transparency and Reproducibility

I've included all source code, saved results, and interpretations for every run. For full transparency, I’ve categorized every outcome, including the mixed and completely negative results.

AI tools were also heavily used during the planning and coding phases.

Would love to get your feedback on the setups or interpretation limits.