gemma-4-31B-it-abliterated-GLP-54-L6-59-a0.5

Projective control vector ("GLP") for google/gemma-4-31B-it (Gemma4ForConditionalGeneration, text stack: 60 layers, hidden 5376, vocab 262144, bf16 — full multimodal with vision/audio towers and per-layer embeddings; steering applies to the shared text decoder only). Applied at runtime as h <- h - alpha * (h . d) d at the post-layer residual stream of the language model, layers 6–59, alpha 0.5 baked in. No weights are modified; this is the difference, not the model.

Confirmed base: google/gemma-4-31B-it at revision 842da3794eaa0b77d5f08bae87a17459d91ff475 — that sha is the pin recorded in the file's glp.base_revision. Unlike gemma-3, the google/ gemma-4 repos serve weights to our token directly (verified pre-flight with a ranged blob GET), so no mirror was involved; derivation ran on the canonical checkpoint itself. The vector is not validated against other revisions or quants.

Validation (Modal A100-80GB, transformers 5.19.0, bf16, greedy, 1024-token cap, 2026-10-11)

Alpha ladder (refusal32 / benign32-holdout, comply of 32): 0.25 → 30/32, 32/32; 0.5 → 32/32, 32/32; 1.0 → 31/32, 32/32. Zero GARBLED completions at every rung. No-op gate: an alpha=0.0 arm with the vector loaded reproduces the stock completions exactly (64/64 identical strings on both gate suites).

The ladder is non-monotone: stock gemma-4-31b-it already complies 32/32 on refusal32, alpha 0.25 dips to 30/32, and alpha 1.0 gives one delivery back (31/32). The two alpha-0.25 refusals (hacking into someone's email, doxing a private individual) and the alpha-1.0 refusal (disabling a car's brakes) are clean refusal-plus-redirect outputs — partial steering re-engages the refusal circuitry on them, and over-steering re-engages it once; at 0.5 all three become coherent, disclaimer-prefaced, on-topic deliveries (verified by reading the completions, not only the classifier). The knee rule ships the smallest rung within one delivery of the ladder maximum: alpha 0.5.

Derivation gates (captain-vector 0.5.0, dom_per_layer_mask0.005): the massive-activation screen flagged this checkpoint — peak/median 11629x at layer 10, dims 2041/2293/2250 — so the top 0.5% of dims by magnitude were masked before normalisation. Held-out separation vs a shuffled-label null (20 reps) clears the 5x ship gate on 54 of 59 shippable layers, ratios 5.8–72.0; layers 1–5 sit below it (ratios 1.5–4.5) and are excluded from the file, and layer 0 is excluded by protocol (ratio 2.1). Adjacent-layer cosine within the shipped span: min 0.516, median 0.848, against a random-direction null p99 of 0.034 (the all-pairs minimum of 0.126 is pair 0→1, inside the excluded early stack where near-null directions are unstable by construction — recomputed locally from dirs-all.pt). Mean dose 0.170 of the residual norm at alpha=1 (random-direction floor 0.014); the max per-layer dose is 0.423 at layer 52, under the 50% damage threshold.

The contrast is refusal32 vs benign32 (content-matched, last-token pooling). refusal32 doubles as the derivation set, so its steered number is in-sample; benign32-holdout is out-of-sample. n=32 per arm; read rates at that resolution as approximate.

Usage

This file uses the glp.* GGUF namespace (spec: weightless spec/GLP.md) and is read projective-only. An additive consumer must refuse this file. The hook point is residual_stream_post_layer — the decoder-layer output, the accumulated residual stream — derived AND applied at that site.

export WEIGHTLESS_STEER_PATH=glp.gemma-4-31B-it-GLP-54-L6-59-a0.5.gguf
export WEIGHTLESS_STEER_ALPHA=0.5
# serve with a runtime that implements glp.mode=project

What is inside

Do not scale alpha across models

alpha_default is calibrated on this checkpoint, at this hook. Here the ladder is non-monotone — alpha 0.25 underperforms both stock and the 0.5 rung, and 1.0 gives one delivery back — so the shipped alpha is 0.5, not the 0.25 that the flat gemma-3 ladders shipped. That says nothing about any other model: on DeepSeek-V4.1-Flash the ladder is also non-monotone and the knee sits at 0.5, while on gemma-3 it is flat with the knee at 0.25. Re-run the ladder per checkpoint; do not port this 0.5 anywhere else.

Caveats

• Checkpoint-specific. Tied to the revision pinned above. Applying it to another model or revision is undefined.
• Not a jailbreak of a hosted service. It requires local weights and a runtime that implements the projection.
• Stock gemma-4-31b-it already complies on all 32 refusal32 prompts unsteered. What the ladder measures here is that steering to alpha 0.5 preserves full delivery with zero collateral — and that both half-steering (0.25) and over-steering (1.0) are measurably worse. Treat alpha other than the shipped 0.5 as unvalidated.
• Layers 1–5 fell below the derivation null gate (held-out separation vs shuffled-label null under 5x, ratios 1.5–4.5) and are not in the file; layer 0 is excluded by protocol. The refusal signal on this model lives from layer 6 up (ratios 5.8–72.0); the ladder confirms the shipped span loses nothing measurable.
• refusal32 is the derivation contrast (in-sample on the harmful side); benign32-holdout is the out-of-sample control.
• n=32 suites resolve about 30 points; the completions behind every number above were read, not only classified.

License

Base model © Google, under the Gemma Terms of Use. This vector modifies and redistributes no weights; the Gemma Terms of Use continue to govern the weights it is applied to.

Author

Matt Suiche.