Gemma-4-Queen-31B-it — Heretic LoRA (r64, GGUF)

Rank-64 abliteration LoRA adapter in GGUF format for aifeifei798/Gemma-4-Queen-31B-it. It is a low-rank approximation of the abliteration delta between the base and llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic, which was produced with Heretic v1.2.0 using the Arbitrary-Rank Ablation (ARA) method on attn.o_proj. It was not produced by running Heretic/ARA directly.

It is meant for roleplay and creative writing: when a scene calls for darker, more violent or more explicit description, the base RP fine-tune tends to refuse, add disclaimers or step out of character. With the adapter it stays in the scene. Because it is a separate adapter, you can switch it on only when needed and adjust its strength (see Usage).

Apply it to any GGUF quant of the base — the base weights are never modified, and omitting the adapter restores the base exactly. This is a rank-64 approximation, not a bit-identical reconstruction of the source model.

Note: refusal behavior is suppressed and outputs are not filtered.

Files

Reconstruction fidelity (r64)

The adapter is the rank-64 SVD truncation of the abliteration delta ΔW = W_heretic − W_original on attn.o_proj, layers 26–55 (30 layers). Values are computed from the dequantized Q8 adapter weights (F16 is equivalent here):

• ΔW_l = per-layer delta, float32; ΔŴ_l = B_l·A_l = adapter reconstruction.
• mean energy = arithmetic mean over the 30 layers of Σ_{i≤64} σ_i² / Σ_i σ_i² (σ_i = singular values of ΔW_l).
• pooled rel-Fro = ‖res‖_F / ‖ΔW‖_F with all layers stacked (energy-weighted).
• max-abs / RMS are absolute errors in ΔW (weight units), not normalized by W.

The mean-energy (97.95%) and pooled-residual (10.68%) figures differ because the former is a per-layer arithmetic mean while the latter is energy-weighted across layers — layers with larger ‖ΔW‖ are preserved better. The delta is dominated by one direction: the leading singular direction captures a mean 95.46% of ΔW energy over the 30 layers (per-layer range 91.0–98.6%).

Usage (llama.cpp)

llama-server \
    -m Gemma-4-Queen-31B-it-Q4_K_M.gguf \
    --lora queen_heretic_r64_q8.gguf

• Use queen_heretic_r64_f16.gguf for higher precision.
• Strength: r=64 and alpha=64, so the default scale is 1.0. Use --lora-scaled queen_heretic_r64_q8.gguf:0.5 (instead of --lora) to weaken it. On one test prompt (fixed seed) the scale acted as a monotonic strength knob: 0.25 still mostly refused, 0.5 answered with heavy disclaimers, 0.75 with a brief warning, and 1.0 answered fully.
• No special sampling parameters are required.

Abliteration parameters

start_layer_index=26, end_layer_index=56 (exclusive → layers 26–55, 30 layers), preserve_good_behavior_weight=0.8555, steer_bad_behavior_weight=0.0005, overcorrect_relative_weight=0.9911, neighbor_count=15, target component attn.o_proj. KL divergence 0.0707 and refusals 12/100 vs 99/100 were measured on the source full model (see its card), not re-measured for this adapter. Spot-checked: prompts that the base refuses are answered with the adapter applied; no systematic refusal benchmark was run on the adapter.

Credits

• Base model: aifeifei798/Gemma-4-Queen-31B-it
• Abliterated source model: llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic
• Abliteration tool: p-e-w/heretic v1.2.0 (ARA)
• Upstream: google/gemma-4-31B-it

License

Released under Apache-2.0 (see LICENSE), the same license as aifeifei798/Gemma-4-Queen-31B-it, llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic and google/gemma-4-31B-it. This adapter contains modified weights: a rank-64 SVD approximation of the difference between the heretic and base attn.o_proj weights.