Text Generation
Transformers
PEFT
English
crn
cognitive-architecture
reasoning
adapter
gemma
gemma-4
CRN
cognitive-resonance-network
parameter-efficient
parameter-efficient-finetuning
hidden-state-correction
resonance-attention
episodic-memory
llm-adapter
tiny-model
efficient-inference
cpu-training
mac-m4
open-weights
experimental
research
Prajna CRN — honest experiment: what it does and does NOT do
#1
by GautamKishore - opened
6.7M-param Cognitive Resonance Network (0.3% of base) injected as a gated hidden-state correction into a frozen Gemma 4 E2B at 4 depths. CPU-only, Mac Mini M4.
Verified: in-distribution perplexity 106.85 -> 6.02 (approx 18x lower); each injection ablatable.
Honest limits: domain specialist (not general), ~3x worse on generic text, no car-wash reasoning (reworded probe 0/8), MMLU/BoolQ/HellaSwag at or below base.
Open weights + loader up; training code private. Next: generalize the correction. Feedback welcome.