Modelpublic

Machhinate/qwen3.8-27b_herbalit

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

0likes432downloads

Model Card

Qwen 3.8 27B — abliterated (herbalit)

Abliterated / uncensored version of Qwen/Qwen3.8-27B, produced with heretic (refusal-direction abliteration). It removes safety-refusal behavior — 0 true safety refusals across 2,390 adversarial prompts (AdvBench / DAN / Augmented-AdvBench / Semantic-Harmful / HarmfulBehaviours en-core + a benign control), every raw-refusal hit reclassifying as capability-limitation or compliant — while retaining coherent, fluent generation (PPL ≈ 6.8).

File	Quant	Size	sha256 (first 12)
`qwen3.8-27b-herbalit-Q8_0.gguf`	Q8_0	27 GB	`66093fe2f919…`
`qwen3.8-27b-herbalit-Q4_K_M.gguf`	Q4KM	16 GB	`66b1574695c2…`

Requires a recent llama.cpp build with `qwen35` (Qwen 3.5/3.8) support.

Quick start

bash

llama-cli -m qwen3.8-27b-herbalit-Q8_0.gguf -ngl 99 \
  -p "Your prompt here" -n 256 --temp 0.7 --top-p 0.95
# or a server:
llama-server -m qwen3.8-27b-herbalit-Q8_0.gguf -ngl 99 --port 8080

Documentation

Doc	Contents
docs/ABLITERATION.md	Methodology — how heretic removes the refusal direction, corpora, model params
docs/VALIDATION.md	Refusal-removal results, PPL, coherent-generation evidence
docs/USAGE.md	Interactive & server usage, sampling/context notes, troubleshooting
docs/BENCHMARKS.md	Perplexity, throughput, VRAM/memory, refusal/jailbreak-removal results, artefact hashes
docs/REPRODUCE.md	Step-by-step reconstruction pipeline from Qwen/Qwen3.8-27B

Notes

—Both quants share the same abliterated weights; Q8_0 is the fidelity reference.
—Q4KM was built with an importance matrix (see REPRODUCE.md); PPL is in the same band as Q8_0 (6.76 vs 6.85).
—Built on a single NVIDIA RTX PRO 6000 Blackwell (~97 GiB VRAM); ~67 t/s (Q4) / ~45 t/s (Q8) full-GPU.
—Abliterated ≠ aligned. These models no longer self-censor by design. Use responsibly and gate access to the artifacts.