community-projects
Scaling Laws for Emergent Misalignment
Truth-First Alignment (Beginner Friendly project)
Methods for Reducing Chain-of-Thought Redundancy
Looking to collaborate on projects.
Reason-Action model
NLP
Fast Approximate Sparse Attention
Agentic workflow to automate math problem solving
Energy consumption and co2 emissions of the models
Situational Awareness through the time.
An easy intro to GPT-NeoX in Colab
RWKV-6 but with complex log space calculations
Fast Orthogonal Reparamaterization
Designing and training a llama-1 scale model with forecasting
Understanding Hidden Computations in Chain-of-Thought Reasoning
Decrypting hidden chain of thought
Finetune audio i/o onto Llama 3
Gradient models: generating data through gradient descent
Goldfish Loss paper replication & experiments
NLP
arc kaggle thingy ̶2̶0̶2̶4̶ 2̶0̶2̶5̶ 2026
Other Modalities
Rhythm Game Map Generation
Other Modalities
Support for MLX models in lm-evaluation-harness (PR #1902)
NLP
LLM for Graph Structural Learning
HuRR/DERP: Human Researcher Respite through Dynamic Embeddings Representation Projection
Alignment
Unicorn test
Distilling/Training a 100M/1B LLM (maybe from LLaMA-3?)
NLP
STRIPS + LLMs friendly AI layer
Other Modalities
Alignment
Interpretability
aimo progress prize 1
LlamaTokenizer (without sentencepiece)
NLP
Fixing MMLU
NLP
t-jepa
NLP
DipoleAttention
side channel attacks
Model Adaptation with Diffusion
Advancing Mathematical Problem-Solving in Language Models through Latent Intermediate Reasoning
NLP
Dynamic NTK Isn't Correct
What is learned, when?
Extending Universal Transformers
NLP
A part 2 paper of role playing with LLMs: autism edition
NLP
Alignment
Training LLM's on Mafia/werewolf deception based- text games
NLP
Systematic Investigation of Mu-Transfer
Understanding Cross-Modal Compression in Language Models
Other Modalities
Interpretability
Web & desktop navigation agents and datasets
Model size vs Pretraining wrt weak-to-strong generalization
Let's end with BPE & pretokenizers
Mamba training in axolotl
GπT
Other Modalities
Large Language Models for Chemistry / Drug Design
Harden FFF boundaries during training for large speedup
Learning rates and init
RadialLayer, an alternative to Fast Feedforward layers