ceselder/gemma4-26b-a4b-nla

NLA for gemma-4-26B-A4B-it (layer 20)

Natural Language Autoencoder trained with EasyNLA/nanoNLA: 49.4% held-out FVE (800 GRPO steps; single-vector baseline −9.6%, i.e. all FVE is text-derived).

• av/ — merged bf16 verbalizer (Gemma4ForCausalLM, text-only)
• av_lora_iter800/ — the RL LoRA adapter alone
• critic/ — 24-block reconstructor + value_head.safetensors Injection marker ㈜ (id 246566); sidecars included. Recipe + full story: report gemma4-nla + github.com/asherps/EasyNLA/pull/1.