LoRA adapter for an Activation Verbalizer (AV) over google/gemma-4-E2B layer-0/embedding
injection (marker token U+3297, injection scale sqrt(1536)). Starting point was the
prior-deviation SFT checkpoint (step 5000); this adapter adds 500 steps of GRPO against a
dense semantic-similarity reward (MiniLM cosine between the generated description and the
activation's source document).
Exact-document top-1 retrieval improved from 3x chance (base) to 0.0121 (7x chance),
worst-case p across 4 negative-sampling seeds 0.0007.
Improvement replicated in two independent training seeds with dose-response
(0 -> 200 -> 500 steps) and transferred to a lexical (tf-idf) metric the reward never
saw.
The OOD stress test (295 never-seen documents, 4 unseen domains) and the blind-judge
comparison were run on the seed-0 checkpoint of this same recipe. Out of distribution a
topic-level ranking signal survives (paired dense-vs-base p = 0.008, in the
semantic-embedding metric only; the lexical metric is flat OOD and top-1 is at chance for
both models). The blind judge preferred the trained model on 28 of 36 decided pairs of 100
(64 ties; p = 6e-4, tie-conservative p = 0.028). This seed-1 adapter shares the training
recipe; its own held-out numbers above are seed-1-specific.
Load with PEFT on top of google/gemma-4-E2B, inject the activation at the ㊗ marker
position in the input embeddings, normalized to L2 norm 39.19. Trained and evaluated
entirely on consumer hardware (4 GB laptop GPU + two used Tesla M40s).
