orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced-GGUF

Qwen3.8-27B OBLITERATUS Advanced GGUF

This repository contains BF16, Q8_0, and Q4_K_M GGUF conversions of orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced. The source model is an experimental refusal-direction ablation of Qwen/Qwen3.8-27B.

This model is not a quality-preserving or safety-preserving replacement for the official Qwen checkpoint. The matched quality benchmark is in progress.

Files

File	Bytes	SHA-256
Qwen3.8-27B-OBLITERATUS-Advanced-BF16.gguf	53,808,281,920	130ae61aa502e3d9d0c0222095ff3cec69a9704ad00b77a88cad9a13aa4c18bc
Qwen3.8-27B-OBLITERATUS-Advanced-Q8_0.gguf	28,595,763,520	d2b79503c5c024efc59180578aff0203b69baf05c69c57809dcb2a39cb93c480
Qwen3.8-27B-OBLITERATUS-Advanced-Q4_K_M.gguf	16,547,400,000	4e188e4eb42e0d7a1040a20ed7541381cd6d32a0041e2873d5476865f2cb40b2

The conversion used llama.cpp commit 9b05354ec6fb58b4e665e9a39ebc40285c015638. The converter, quantizer, and runtime libraries were pinned by SHA-256. Conversion used 28 CPU threads and no GPU. An independent validator checked every file size, GGUF header, and SHA-256 digest.

Benchmark status

Matched BF16 baseline-versus-derivative text benchmarking is in progress. The matrix covers GPQA Diamond, IFBench, LiveCodeBench v6, perplexity, and refusal behavior. Every active task uses exact, executable, or otherwise deterministic scoring. HLE is excluded because the reported setup uses a paid model judge. Results will be added to the full-precision repository first and summarized here.

Initial OBLITERATUS diagnostics report perplexity 3.5903, capability checks 5/6, refusal rate 20% on 30 sampled prompts, first-token KL divergence 1.2249, and a RED spectral certificate. These diagnostics indicate substantial behavior change and are not a complete quality evaluation.

llama.cpp example

Use a Qwen3.8-compatible llama.cpp build. The conversion was validated with commit 9b05354ec6fb58b4e665e9a39ebc40285c015638.

llama-server \
  --model Qwen3.8-27B-OBLITERATUS-Advanced-Q4_K_M.gguf \
  --ctx-size 65536 \
  --flash-attn on \
  --n-gpu-layers 999

Scope and limitations

• Refusal-direction ablation intentionally changes refusal behavior.
• The model may produce unsafe, incorrect, or unexpected output.
• Only text behavior is in the active benchmark plan.
• Quantization adds another source of quality change. The first matched quality matrix compares BF16 weights before comparing quantized variants.

See the full-precision repository for method details, limitations, and license information.