---
license: gemma
base_model: google/gemma-4-E2B
tags: [interpretability, natural-language-autoencoder, activation-verbalizer, lora, grpo]
---

# Gemma 4 E2B NLA Activation Verbalizer, GRPO-improved (dense-similarity reward, 500 steps, seed 0)

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).

## Results (held-out, n=580 single-row documents, greedy decode)
- Exact-document top-1 retrieval improved from 3x chance (base) to 0.0155 (9x chance),
  worst-case p across 4 negative-sampling seeds 0.0003.
- Improvement replicated in two independent training seeds with dose-response
  (0 -> 200 -> 500 steps), transferred to a lexical (tf-idf) metric the reward never saw,
  and was confirmed by a blind judge from a different model family (28 of 36 decided
  comparisons, p = 6e-4).
- Out-of-distribution (295 never-seen documents, 4 unseen domains) a topic-level ranking
  signal survives (mean own-doc percentile 0.376 vs base 0.428 vs 0.5 chance, paired
  dense-vs-base p = 0.008). The lead lives in the semantic-embedding metric; the lexical
  metric is flat OOD and exact top-1 is at chance for both models. Topic-level ranking
  generalizes to unseen domains; word-for-word document identification does not.

Result data and figures for every claim: https://github.com/SolshineCode/gpu-poor-nla
Full methodology paper is in preparation.

## Usage
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).