SRT-NLA gemma-4-31B-it artifacts (L47)

The retrieval indexes that let a frozen gemma-4-31B-it turn its L47 states into full sentences, with zero training.

Layer 47 is the backbone's cross-modal alignment peak. Encode any L47 state — a text's last token, or the mean over an image's soft-tokens — and these indexes decode it by nearest-neighbour retrieval (github.com/space-bacon/SRT, paper_nla.md §11.6–§11.7).

On "peak". That ranking is measured with the raw-cosine read-out these indexes use. A fitted linear probe orders the layers differently and puts L20 ahead of L47 (R@1 0.293 against 0.230 on 4,000 held-out COCO pairs). The depth profile is a joint property of the backbone and the probe, so the two orderings are both real and answer different questions. Full nine-layer tables, scaling curves and scripts: RiverRider/srt-depth-probe-artifacts. L47 remains the tap these artifacts are built and validated at.

What they power

• Open-vocabulary image captioning: 5/5 CIFAR natural images retrieve on-topic captions at rank 1 from the 10k pool (paper §11.6.3), live in the SRT-Sunstone demo.
• Text-state read-out at centered fve 0.695 (NN) inside a calibrated anchor frame (replay 0.994 / floor 0.494).
• Retrieval is the primary decoder on this backbone; the generative companion for sampled candidates is RiverRider/srt-nla-av-gemma4 (best-of-K with oracle rerank).

Files

Usage (retrieval decode)

import torch, torch.nn.functional as F
from huggingface_hub import hf_hub_download

idx = torch.load(hf_hub_download("RiverRider/srt-nla-gemma4-artifacts",
                                 "caption_index_L47.pt", repo_type="dataset"),
                 map_location="cpu", weights_only=False)
pool, texts = idx["vectors"].float(), idx["texts"]
mu = pool.mean(0)
pc = F.normalize(pool - mu, dim=-1)
# v: your L47 query state (text last-token, or mean over image soft-tokens,
# centered by a mean of the SAME modality — see paper §11.6.3)
sims = pc @ F.normalize(v - mu_query_modality, dim=-1)
print(texts[int(sims.argmax())])

Protocol notes: gemma-4 must be loaded as Gemma4ForConditionalGeneration (not CausalLM), and every text re-encode must prepend BOS.