Fuck it I'm gonna cook Gemma 4 31B with de-prose and de-euphemism. E4B results were good enough (though I went overboard and made the model write like a middle schooler, need to tone it down). This is my plan.
60 ablation trials.
Optimizer: two finetuned BERT classifiers, one for purple axis, one for euphemism axis.
Guardrails: repetition detectors (intra-reply, structural, phrase detection, etc.), perplexity vs human writing text, gen perplexity vs base text.
Sampler: TPE instead of gradient descent or Bayesian because I punish brain damage and cheating hard and the deltas in final scores will be huge for cheating attempts, TPE just discards these fuck-ups entirely instead of trying calibrate on them.
Flooring: babi benchmark (it's state tracking so it's relevant for RP) -> Take best 20% trials, average their scores -> add 20% and get acceptable floor -> run benchmark on all passing trials -> keep the best 10 and eyeball their outputs
>what is babi
{"id": "babi_t5_4", "system": "Read the statements, then answer the question with a single word.", "user_turns": ["Mary travelled to the garden.\nMary journeyed to the kitchen.\nBill went back to the office.\nBill journeyed to the hallway.\nJeff went back to the bedroom.\nFred moved to the hallway.\nBill moved to the bathroom.\nJeff went back to the garden.\nJeff went back to the kitchen.\nFred went back to the garden.\nMary got the football there.\nMary handed the football to Jeff.\nWhat did Mary give to Jeff?"], "checks": [{"expect": ["football"]}]}
I'm low on vast credit so might run out before the run is done. Hope vast keeps my hard disk data for some time if I run out of money.
