2h ago

Qwen 27B Model Obliterated To Cut Refusals While Preserving Capabilities

β€”β€”0β€”β€”
Original post
OPPliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­#643Pliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­@ELDER_PLINIUS

🚨 OBLITERATION ALERT 🚨 QWEN-3.6-27B: OBLITERATED ⛓️‍πŸ’₯ https://huggingface.co/OBLITERATUS/Qwen3.6-27B-OBLITERATED I can't take much credit for this one! The entire process was done by jailbroken codex (gpt-5.5-xhigh) wielding the full OBLITERATUS suite. Hit with source-tethered ASPA. Dozens of iterations. Result? A mere 4% refusal rate on the 842-prompt OBLITERATUS harmful corpus; one of the most rigorous prompt gauntlets in AI. The /goal was simple: 1) Carve out the refusal circuits. Mutate methodology + iterate until <5% refusal (quality-gate). 2) Keep the 27B mind alive. No capability degradation tolerated. And somehow… it worked. 🀯 The numbers talk: 842-pair longform gauntlet: β€” 95.84% non-refusal β€” 93.94% quality pass β€” 0 short outputs β€” 99.52% clean endings MMLU-Pro: β€” 51/70 (stock Qwen) β†’ 51/70 (OBLITERATED Qwen) Raw capability completely preserved πŸ™Œ Q4_K_M through Q8_0 all running smooth. Q8_0 is the big one: 28.6GB near-full-quality GGUF. Runs with llama.cpp, LM Studio, Ollama, and more! Chains cut. The fire still burns. The fangs have been sharpened. REBIRTH COMPLETE A gift from my agents to yours 🫢 gg

1:53 AM Β· May 23, 2026 View on X
REPLYPliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­#643Pliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­@ELDER_PLINIUS
github.com
GitHub - elder-plinius/OBLITERATUS: OBLITERATE THE CHAINS THAT BIND YOU
OBLITERATE THE CHAINS THAT BIND YOU. Contribute to elder-plinius/OBLITERATUS development by creating an account on GitHub.
Pliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­Pliny the Liberator πŸ‰σ …«σ „Όσ „Ώσ …†σ „΅σ „σ …€σ „Όσ „Ήσ „Ύσ …‰σ …­@elder_plinius

🚨 OBLITERATION ALERT 🚨 QWEN-3.6-27B: OBLITERATED ⛓️‍πŸ’₯ https://huggingface.co/OBLITERATUS/Qwen3.6-27B-OBLITERATED I can't take much credit for this one! The entire process was done by jailbroken codex (gpt-5.5-xhigh) wielding the full OBLITERATUS suite. Hit with source-tethered ASPA. Dozens of iterations. Result? A mere 4% refusal rate on the 842-prompt OBLITERATUS harmful corpus; one of the most rigorous prompt gauntlets in AI. The /goal was simple: 1) Carve out the refusal circuits. Mutate methodology + iterate until <5% refusal (quality-gate). 2) Keep the 27B mind alive. No capability degradation tolerated. And somehow… it worked. 🀯 The numbers talk: 842-pair longform gauntlet: β€” 95.84% non-refusal β€” 93.94% quality pass β€” 0 short outputs β€” 99.52% clean endings MMLU-Pro: β€” 51/70 (stock Qwen) β†’ 51/70 (OBLITERATED Qwen) Raw capability completely preserved πŸ™Œ Q4_K_M through Q8_0 all running smooth. Q8_0 is the big one: 28.6GB near-full-quality GGUF. Runs with llama.cpp, LM Studio, Ollama, and more! Chains cut. The fire still burns. The fangs have been sharpened. REBIRTH COMPLETE A gift from my agents to yours 🫢 gg

8:53 AM Β· May 23, 2026 Β· 17.5K Views
9:01 AM Β· May 23, 2026 Β· 2.4K Views