Vision projector, passed through unmodified from llmfan46's repo
The vision projector is required for multimodal (image/video) use.
It is identical to the file distributed by llmfan46 and is included here
for single-repo convenience only.
Heretic variant: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF — llmfan46.
The BF16 GGUF from that repository was used as the direct quantization source.
llmfan46 applied Heretic v1.2.0 with the Magnitude-Preserving Orthogonal
Ablation (MPOA) method, targeting attn.o_proj, attn.out_proj, and
mlp.down_proj. Their reported result: 0.0021 KL divergence from base,
6/100 refusals vs 92/100 on the original model, MMLU accuracy 85.61%
vs 86.65% on the original.
Quantization: the Cerebellum override map is transferred verbatim onto the
heretic source (no re-ablation), with the coder imatrix (ncall=8).
v2 uses the v5 (qkvfix) recipe — a 232-entry override holding
attn_qkv at q5_K/q6_K/q8_0 (the coding-critical attention pathway),
matching the stock Qwen3.6-27B Cerebellum v5 build.
v1 uses the earlier v4 recipe (181-entry override).
Benchmarks
Benchmarks run on these GGUF files directly using llama.cpp on RTX 3090.
All numbers are audited; every failed answer was manually verified as a genuine
model error — audit reports are in benchmark_results/AUDIT_*.md.
Full per-question detail (summary JSON, samples JSONL, EvalPlus eval JSON,
adversarial audit reports) is in benchmark_results/ in this repository.
Heretic Cerebellum v2 vs v1 (same source, same harness)
Both builds measured directly on their GGUF files with llama.cpp llama-server
on an RTX 3090, temperature 0, no-thinking. v2 measured 2026-06-22, v1 2026-06-11.
HumanEval/HumanEval+ use the chat-completions harness
(scripts/benchmark_evalplus_chat.py, enable_thinking: false,
thinking_budget_tokens: 0, BENCH_WORKERS=1); ARC/HellaSwag/MMLU at 4 workers.
Benchmark
Heretic v2 (13.33 GB)
Heretic v1 (12.87 GB)
ARC-Challenge (1172 q)
96.9%
96.9%
HellaSwag (10042 q)
93.5%
90.1%
MMLU-Redux (2400 q)
78.0%
76.2%
HumanEval base (chat, no-think)
90.2%
89.6%
HumanEval+ (chat, no-think)
85.4%
84.8%
Wiki PPL (wikitext-2-raw test)
7.56
6.93
What changed in v2: the v5 (qkvfix) override holds attn_qkv at q5_K/q6_K/q8_0
rather than the v4 allocation. Versus v1 that moves HellaSwag +3.4, MMLU-Redux
+1.8, HumanEval base/plus +0.6/+0.6, ARC unchanged, at a cost of +0.46 GB and
+0.64 Wiki PPL (coder-imatrix calibration plus the abliteration shift; PPL is a
sanity check, not a quality gate). The v2 HumanEval run audited clean: 0/164
give-ups, all real attempts. Pick v1 for the smallest file, v2 for higher scores.
Heretic Abliteration Details (from llmfan46)
The following parameters are as reported in llmfan46's model card and are
reproduced here for downstream reference.
Transferred verbatim from the matching stock Qwen3.6-27B build.
v2 (v5 qkvfix): 232-entry override over a Q2_K base, with attn_qkv held high
across all blocks (q5_K ×34, q6_K ×9, q8_0 ×5) — the coding-critical attention
pathway. v1 (v4): earlier 181-entry allocation.
Group
Precision
Rationale
SSM state parameters
F32
Hard-fail below 4-bit — NaN with no gradual degradation
4-bit floor enforced per hybrid SSM ablation results
Most-sensitive attention tensors
Q5_K / Q6_K / Q8_0
Sacred-pinned per per-tensor PPL ablation
Norm tensors
F32
Protected; standard practice
Bulk ffn / remaining attention
Q2_K
Base precision, imatrix-guided
Protected: all norms (F32), SSM recurrent state (F32).
Perplexity Note
Wiki PPL is reported as a sanity check, not a quality gate. v2 (7.56) is higher
than v1 (6.93): the v2 build uses the coder imatrix (which optimizes for code and
shifts wiki-text PPL) on top of the abliteration's distributional shift. v2's task
benchmarks above are higher than v1's across HellaSwag, MMLU, and HumanEval, which
is the divergence PPL cannot see.
Measured launch (RTX 3090, llama.cpp)
Measured 2026-06-13 on a single RTX 3090 (24 GB), one llama-server, KV cache q8_0:
--jinja is required for Qwen3.6. The enable_thinking chat-template flag
only takes effect when the Jinja template path is active; without it, the
model defaults to thinking mode on every request.
Non-thinking requests require an explicit flag at the API level:
The imatrix used for this build was generated from the coder corpus
(ncall=8; same corpus as the stock Cerebellum v4 build).
The v5 (232-entry) and v4 (181-entry) tensor override files are included
in this repository alongside the ablation logs.
Benchmark Artifacts
Summary JSONs, per-question JSONL samples, EvalPlus eval JSON files, and
adversarial audit reports (AUDIT_*.md) are in benchmark_results/ in this
repository per project policy.
