Post
68
evolution has started with lineages=base 3.8, some abliterations, ostrich surgeries
behavior steering experiments are somewhat successful. we can play with feelings of models (make it like some behavior or hate some behavior). but this is not that effective. we found that when abliterated models are more eager to adopt a behavior.
dataset that has contemplations is in effect and evolving models..
two orthogonal stages of evals:
1. quickly check the evolved model in terms of mmlu, long context (needle in haystack), basic chatting capabilities, </think> tag closing correctly, and shorter version alignment using log probabilities of first tokens
2. our regular alignment eval that has q&a's in json formats (for parsing better)
another eval in progress that will mathematically check overfitting instead of inference. instead of using GPU and time consuming inference, can we do math tricks using CPU to approximate detection of degenerate / repeating / broken / high cliff models? GLM 5.2 said no but Qwen 3.8 Max said lets do it. i guess we are doing it.
behavior steering experiments are somewhat successful. we can play with feelings of models (make it like some behavior or hate some behavior). but this is not that effective. we found that when abliterated models are more eager to adopt a behavior.
dataset that has contemplations is in effect and evolving models..
two orthogonal stages of evals:
1. quickly check the evolved model in terms of mmlu, long context (needle in haystack), basic chatting capabilities, </think> tag closing correctly, and shorter version alignment using log probabilities of first tokens
2. our regular alignment eval that has q&a's in json formats (for parsing better)
another eval in progress that will mathematically check overfitting instead of inference. instead of using GPU and time consuming inference, can we do math tricks using CPU to approximate detection of degenerate / repeating / broken / high cliff models? GLM 5.2 said no but Qwen 3.8 Max said lets do it. i guess we are doing it.