Downloads · 30 days
118
6% of all-time downloads
orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced-GGUF
Qwen3.8-27B-OBLITERATUS-Advanced-GGUF is a text generation model from orwelian84. Use it when you need the model to write or continue text. It is set up for gguf. The card lists the license as apache-2.0.
This repository contains BF16, Q80, and Q4KM GGUF conversions of orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced. The source model is an experimental refusal-direction ablation of Qwen/Qwen3.8-27B.
Downloads · 30 days
118
6% of all-time downloads
All-time downloads
2.1K
Public
Repo size
99 GB
Likes
1
Public
Click a slice to open those files.
.gguf99 GB · 100%
From the Hugging Face model README
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.
| 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.
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.
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
See the full-precision repository for method details, limitations, and license information.