eric-tramel/audiolens-gemma-4-E2B-it

Audiolens runtime lenses

Runtime Jacobian lenses for google/gemma-4-E2B-it. The base model remains governed by the upstream Gemma model card and Apache-2.0 license. Fitting and bundling code comes from audiolens under MIT; the estimator comes from anthropics/jacobian-lens under Apache-2.0. Lens files in this repository are offered under cc-by-sa-4.0.

Recommended boundary lens

gemma-4-E2B-it-crossmodal-6f55ad1f2d85-b2000.pt is the default lens for Moodmic. Lens B is fitted over one deterministically shuffled 2,000-prompt corpus: 1,000 pinned LibriSpeech utterances and 1,000 licensed FSD50K event clips. Every prompt contains the complete audio/instruction/assistant context, but B targets the label-free assistant response boundary. A causal sentinel verified that later answer content has exact-zero forward and gradient influence on that target.

The lens maps fit-valid audio-position residuals from L0–L13 to target L34. Moodmic reads L13, the final audio-position layer with a nonzero causal route to a later response position before Gemma 4's late K/V reuse. The runtime artifact is fp16, 66,065,229 bytes, and has SHA-256 c48a53e90d256aaebeeb1ca7607ab08bc47574ebe7e18f24acd68e2c0ce4e673. See audiolens-crossmodal-boundary-b-run.json for sanitized corpus, fit, runtime, identity, and checksum bindings.

This is an exploratory last-cross-position-interface lens. The frozen model capability gate ended inconclusive_model_lacks_nonspeech_capability; the artifact does not establish an L13–L31 regional J-space, formal J-space component, mood classifier, or general event-recognition capability.

Source-bound audio lens

gemma-4-E2B-it-audio-ee7cd4e42991-audio1000.pt is the prior waveform-only, source-bound audio lens. It was fitted over 1,000 pinned LibriSpeech utterances: 500 clean/train.360 and 500 other/train.500, one 2–4 second native-mono 16 kHz clip per globally unique speaker. Transcripts were retained only as provenance and were never passed to its estimator.

That run fitted L0–L33 to target L34 with skip_first=16. Its fp16 runtime artifact is 160,439,813 bytes with SHA-256 da0ccabf1ee14e4df060f97f31cf0132a0d3f6ed2cb45b6c77738693bc8f1aa9. See audiolens-audio-only-run.json for sanitized source, corpus, fit, checkpoint, stability, and checksum bindings.

Historical mixed lenses

The earlier mixed experiment combines 400 WikiText prompts with 128 processor-validated LibriSpeech utterances. Its text and mixed lenses remain available for reproducibility; audiolens-run.json contains their sanitized completed-run metadata.

Files

• gemma-4-E2B-it-crossmodal-6f55ad1f2d85-b2000.pt — recommended boundary lens; 66,065,229 bytes; SHA-256 c48a53e90d256aaebeeb1ca7607ab08bc47574ebe7e18f24acd68e2c0ce4e673
• audiolens-crossmodal-boundary-b-run.json — sanitized completed-run metadata for Lens B
• gemma-4-E2B-it-audio-ee7cd4e42991-audio1000.pt — source-bound audio-only lens; 160,439,813 bytes; SHA-256 da0ccabf1ee14e4df060f97f31cf0132a0d3f6ed2cb45b6c77738693bc8f1aa9
• audiolens-audio-only-run.json — sanitized completed-run metadata for the source-bound audio lens
• gemma-4-E2B-it-mixed-a84c3b9df539-mixed528.pt — historical mixed528 lens; SHA-256 1beff8be75b67169ff968349ccd953b712dd8f90097e23cb0b1cd2515ee39cad
• gemma-4-E2B-it-mixed-a84c3b9df539-text400.pt — historical text400 lens; SHA-256 f42ffa4c940da6f0b06426ba8e09382c348c652979aae34f32e3b88acae50b48
• audiolens-run.json — sanitized completed-run metadata for the historical mixed experiment

Pin an immutable repository revision and verify the expected SHA-256 before loading any lens.