Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.49
attn.o_proj.max_weight_position	20.30
attn.o_proj.min_weight	1.24
attn.o_proj.min_weight_distance	15.39
mlp.down_proj.max_weight	0.80
mlp.down_proj.max_weight_position	20.14
mlp.down_proj.min_weight	0.80
mlp.down_proj.min_weight_distance	3.73

Performance

Metric	This model	Original model (electroglyph/gemma-3-4b-it-unslop-GRPO-v3)
KL divergence	0.92	0 (by definition)
Refusals	13/100	94/100

Gemma 3 4b unslop experiment v3

An unslop finetune of google/gemma-3-4b-it

Temperature during training was at 1.0 this time around, model is a lot less weird
Rewards changed a little bit. I allowed a small number of sentences with 4+ commas instead of penalizing them all. This has cut down on the number of paranthetical phrases without completely eliminating them.
Lexical diversity score is a bit fancier this time. First I calculated MTLD for 600+ books I have and looked at the mean score. It was almost exactly 100.0, so that's the baseline I aimed for. MTLD of 80-120 all received full points (to avoid too much GRPO chaos), but deviations further than that get increasingly penalized.
I've uploaded a UD-Q4_K_XL GGUF with settings that I grabbed from Unsloth's quant using my lil utility: quant_clone

Basically the same as last time plus the minor changes above.

training code: train.py