there is a new version of ostrich, more successful with long context jobs.
https://huggingface.co/etemiz/Ostrich-27B-260721
apparently when i push for higher AHA score it ends up overfitting and merging it with previous 3.5 version healed the overfittings.
this gave me another idea: what if we merge all the fine tuners like fine tunes going for abliteration (uncensored) and also separately agentic coding. since the merging heals, everybody's overfittings can cancel each other.
if that works, this could mean models that we create are like our overfitted ideologies. we humans all are like knowing and believing some stuff religiously and when or if we can get together those extremes kind of relaxed. socialization of humans look exactly like merging of LLMs..
https://huggingface.co/etemiz/Ostrich-27B-260721
apparently when i push for higher AHA score it ends up overfitting and merging it with previous 3.5 version healed the overfittings.
this gave me another idea: what if we merge all the fine tuners like fine tunes going for abliteration (uncensored) and also separately agentic coding. since the merging heals, everybody's overfittings can cancel each other.
if that works, this could mean models that we create are like our overfitted ideologies. we humans all are like knowing and believing some stuff religiously and when or if we can get together those extremes kind of relaxed. socialization of humans look exactly like merging of LLMs..