Reacts weird
I tried this model in Q4_K_M and IQ4_N_L quantization. In both cases, it makes a lot of mistakes. The Version from mlabonne is much more stable and also uncensored.
True, I will be choosing and updating it with a different iteration with lower KL Divergence but maybe 1-15/100 refusal.
I'll test it and give you feedback.
Did you already upload the update? Because model-00001-of-00012.safetensors is still old.
I have updated the model, with 0.07 divergence, it should be slightly better than one by mlabonne. His model had KL divergence of : 0.08 and 6/100 refusals, (lower the divergence, lesser the harm to model due to ablation)
I have also published another v2 version with even lower divergence at cost of some more refusal here: coder3101/gemma-3-27b-it-heretic-v2