qwen3.8-27b-uncensored-properly-distill-experimental

properly uncensored qwen3.8-27b, poorly tailored for persistent agent and swarm workflows

---

don't use this seriously: it's not that good

believe it to serve as a good demonstration of behavioural tuning possibilities without leading to coherence, ability, or performance degradation

censorship removal isn't just one direction, there's a lot more to how the assistant is shaped which shapes behavioral patterns, reasoning, and responses; heretic ara is a good direction but believe mine to be significantly better

distill is fable, sol, mythos, opus5, and kimi k3 agentic traces (with reasoning), mostly a private collection

this model exhibits a decent amount of behavioural issues because i screwed up the long context traces in data a little, but behaves better in a harness than stock if configured properly (low temp, high top p, low top k, presence and repetition penalties) + harness has protections against looping, early stopping, and long context task tracking attention degradation (arxiv:2601.16746 is good for rollbacks so ctx doesnt end up polluted)

---

influences:

• arxiv:2603.18656
• arxiv:2406.11717
• arxiv:2604.13016
• arxiv:2608.11583
• arxiv:2608.08168
• arxiv:2606.13720
• arxiv:2605.26772
• https://themenonlab.blog/blog/heretic-automatic-abliteration-llm-refusal-removal
• https://www.lesswrong.com/posts/sSyLyc3KDQzboQGWS/thoughts-on-claude-fable-s-silent-safeguards

---

eval *= -1