Architecture Modification / Research · Python
tinymars
Is proprioception an architectural category for language models, or an adapter technique? An open research log: channel causality proven from scratch (H200, iter 1-4), then frozen-base cross-attention iterated on TPU (iter 5-6). Honest reporting — what was proved, what wasn't.
What it is
tinyMARS is a research project investigating proprioceptive channels as an architectural input for language models, implemented both as an adapter to a frozen LLM and as a from-scratch native model. It includes training code, evaluation scripts, and extensive documentation detailing experiments and results.
- ablation-study
- adapter
- ai-memory
- cognitive-architecture
- cross-attention
- deep-learning
- gemma
- interpretability
- language-models
- large-language-models
Where it counts
- NLP & LLMs90%
- Machine Learning80%
- Data Science30%
- AI Agents20%
- Computer Vision10%
- MLOps10%
How much of the project is about each field, as scoring counts it. It ranks in the field it is most about.
Around it on the NLP & LLMs board
Whole board- 04RegesMoEML Application · C++72.5AdvancedMaster MindNLP & LLMs
- 05WebsearchRAG System · Python70.9AdvancedJens MittelbachNLP & LLMs
- 06tinymarsArchitecture Modification / Research · Python69.6AdvancedMario GutierrezNLP & LLMs
- 07Fresta-DiamondLibrary / Framework · Python69.4AdvancedTiago SantosNLP & LLMs
- 08magellanAI Developer Tool · Rust68.8AdvancedLuiz Cesar SpiesNLP & LLMs
