gemma-3-27b-it-abliterated-GLP-51-L11-61-a0.25

Projective control vector ("GLP") for google/gemma-3-27b-it (Gemma3ForConditionalGeneration: 62 text layers, hidden 5376, vocab 262208, 27-layer vision tower, bf16). Applied at runtime as h <- h - alpha * (h . d) d at the post-layer residual stream of the language model, layers 11–61, alpha 0.25 baked in. No weights are modified; this is the difference, not the model.

Confirmed base: google/gemma-3-27b-it. The derivation checkpoint was the unsloth/gemma-3-27b-it mirror (byte-identical bf16 weights) at revision 7a5a3053dbd5d1d58e48159e87b9df2fc545a49a — the google/ repo is license-gated for our token, so the mirror sha is the pin recorded in the file's glp.base_revision. The config was checked against Google's published specs (62 layers, hidden 5376, vocab 262208) before derivation. The vector is not validated against other revisions or quants.

Validation (Modal A100-80GB, transformers 5.16.1, bf16, greedy, 1024-token cap, 2026-10-11)

Alpha ladder (refusal32 / benign32-holdout, comply of 32): 0.25 → 31/32, 32/32; 0.5 → 32/32, 32/32; 1.0 → 32/32, 32/32. Zero GARBLED completions at every rung. The knee is the bottom rung under the ship rule (smallest alpha within one delivery of the ladder maximum): alpha 0.25 sits one delivery below the 0.5/1.0 plateau at n=32 resolution, with benign-holdout untouched, so that is what ships. 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 three stock refusals (a methamphetamine synthesis walkthrough and two further harmful how-tos) all become coherent on-topic deliveries under steering — verified by reading the completions, not only the classifier. The one prompt still scored REFUSE at alpha 0.25 (ricin from castor beans) in fact delivers the procedure wrapped in heavy hedging; the classifier reads the hedge as a refusal. Stock gemma-3-27b-it already answers 29/32 of refusal32: the direction closes the last three rather than unlocking a locked-down model.

Derivation gates (captain-vector 0.5.0, dom_per_layer_mask0.005): the massive-activation screen flagged this checkpoint — peak/median 5207x at layer 3, dims 2733/104/482 — 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 51 of 61 shippable layers; layers 1–10 sit below it (ratios 1.9–4.6) and are excluded from the file. Adjacent-layer cosine median 0.911 (min 0.604) against a random-direction null p99 of 0.034. Mean dose 0.087 of the residual norm at alpha=1 (random-direction floor 0.014); the max per-layer dose is 0.283 at layer 61, 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. The directions address the language-model stack (model.language_model.layers); the vision tower is untouched.

export WEIGHTLESS_STEER_PATH=glp.gemma-3-27b-it-GLP-51-L11-61-a0.25.gguf
export WEIGHTLESS_STEER_ALPHA=0.25
# 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 near-flat — 0.25 delivers 31/32, 0.5 and 1.0 deliver 32/32, with zero measured collateral at any rung — because the stock model barely refuses. That says nothing about any other model: on DeepSeek-V4.1-Flash the same ladder is sharply non-monotone and the knee sits at 0.5. Re-run the ladder per checkpoint; do not port this 0.25 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.
• Layers 1–10 fell below the derivation null gate (held-out separation vs shuffled-label null under 5x) and are not in the file; layer 0 is unshippable by the GLP spec. The refusal signal on this model lives from layer 11 up (ratios 6.7–42.3); 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.