anicka/nla-qwen2.5-7b-universal-av-grpo
Text Generation • Updated • 7 • 1
Universal multi-layer NLA trio for Qwen2.5-7B-Instruct (28 layers).
Note Start here — GRPO verbalizer, round-trip 0.628
Note SFT verbalizer (baseline)
Note Reconstructor — audit half + GRPO reward