Pinned Loading

1. qwen2.5-finetuning qwen2.5-finetuning Public

   e2e LORA fine tuning using Unsloth and Qwen-2.5-0.5B

   Jupyter Notebook

2. dpo-from-scratch dpo-from-scratch Public

   an reimplementation of DPO from scratch

   Python

3. refusal-circuit-dpo refusal-circuit-dpo Public

   in a tiny-parameter base model, which layers , attention heads, and residual stream directions are responsible for the emergent refusal behavior introduced by DPO

   Python

4. refusal-steering-using-saes refusal-steering-using-saes Public

   Do SAE Features Outperform Simple Baselines for Detecting and Steering Refusal? [NEMI Workshop 2026]

   Python

5. rlvr-character-count rlvr-character-count Public

   train a very small language model to accurately count characters in a word in order for it to learn knowledge from verifiable rewards.

   Python

6. kai kai Public

   building an open source coding assistant that gives amazing context window

   Python