ColdSlither Labs

Local LLM research and build. Small models, honest numbers, reproducible recipes. Based in East Orange, NJ. Working toward technical independence, one verified run at a time.

QLoRA fine-tuningRefusal researchQuantizationEval disciplineRTX 5060 Ti 16GB

Capability-preserving uncensoring

0 refusals
OBLITERATUS advanced run on Qwen3-8B with capability intact, served as GGUF. Selection rule: lowest KL divergence under 10 refusals, never min refusal. KL over 0.5 means real damage.

Measured on our prompt set and HF 0/32; harsher shared harness scores 4/10. Both numbers published, neither hidden.

Head-to-head method comparison

Heretic vs OBLITERATUS on the same Qwen3-8B base. Both held 5/5 capability. Refusal: Heretic 30%, OBLITERATUS 40% (n=10). Heretic is the locked default; OBLITERATUS stays as manual fallback.

Anti-collapse training module

7.72 held-out
Arenas residual module, trained and benchmarked on Qwen3-0.6B/1.7B. Beats the naive no-residual foil in every run. Magnitude-normalized residual leads.

Eval discipline

Content-only refusal scoring, 400 to 600 token budgets, matched-budget ablations. A 2x2 isolation proved an apparent serving gap was a token-budget artifact, not a thinking-mode effect. Measure twice, publish once.

What I offer

Research stage. Methods are measured on small harnesses and published with caveats. No enterprise claims, no borrowed logos.