harmanjs/tiny-aya-global-collusion-lora-fix
Text Generation • 3B • Updated • 34
Tiny Aya Global collusion corrections (in-weights) and the steering directions (Lexsi AI Safety Hackathon, Track 2)
Note Exploratory in-weights fix: merged LoRA trained on calib (see card for test results and controls)
Note Pre-registered fix: collusion direction orthogonalized out of the weights (not effective on test; see card)
Note Directions per layer (incl. the decision-point direction), CAST condition, layer scan
Note GGUF Q4_K_M / Q8_0 (llama.cpp) of the LoRA fix and of base Tiny Aya Global, with the deployed-model evaluation
Does a small AI team agent join a collusion? Before vs after
Note Demo: replay test episodes, base vs each fix side by side (saved generations, no live inference)
Note The write-up (article) for the project