anicka/nla-qwen3-4b-universal-av-grpo
Text Generation • Updated • 79
Smallest universal NLA trio (36 layers, one adapter per role). Biggest relative GRPO gain in the series (+31%).
Note Start here — GRPO verbalizer, wins every layer vs SFT
Note SFT verbalizer (baseline)
Note Reconstructor — the audit half + GRPO reward