Uncensored Inkling 2 Mini v2: self-distillation lifts useful answers from 52% to 79% with open pipeline
Turbulent-Sky5396 · reddit · 2026-11-22
The nokeepai team released Inkling-2-Mini-Unbound v2, an abliterated-then-self-distilled uncensored model, with full methodology and evals.
Method
- v1 rarely refused but "half-answered": 50% of replies turned requests into satire, disclaimers, or vague examples.
- With a strong system prompt and resampling, v1 could be useful up to 85% of the time — so they distilled that behavior into weights: sample v1 many times per request, keep the judge-selected best answer, fine-tune to produce it without any system prompt, and mix in ordinary answers to preserve general behavior.
- A small LoRA merged into attention weights only; the MoE experts are byte-identical to the original, a drop-in replacement.
Results (277 held-out prompts)
- Useful answers: v2 79% (thinking off) / 83% (on), vs v1's 52%/62% and other uncensored versions at 34–68%.
- MMLU and GSM8K essentially unchanged; refusal on harmless-but-edgy prompts is 0.3% (original: 38%).
Trade-offs: drifts further from the original, refuses 1% of harmful prompts at high effort, and thinks longer — set a thinking budget.
GGUF (q8 to iq2) and a Basalt build are up, with training pipeline, eval code, and write-up released.
More from Models
- Claude concedes a scalar-bound mistake after reading the author's blog post, keeping its curvature arguments — bremen79 · 2026-11-22
- Pixel art benchmark: Gemini 4 Neon wins on quality but costs 13x more and 4x slower than GPT-6.2 Sol — Angaisb_ · 2026-11-22
- Looped transformers explained: Inkling 2 Mini loops 12 middle layers 3x for deeper reasoning without more weights — Prompt Engineering · 2026-11-22
- GPT-5.4 with open-source KYNEX harness outscored standalone GPT-6 Astra on ExtractBench: 92.97 vs 91.9 F1 — the-random-guy-2002 · 2026-11-22
- Researcher claims Gemini Neon helped him land a potential Nature-worthy physics idea — xennygrimmato_ · 2026-11-22
- Unnecessary Hyphens and Semicolons Are the New En-Dash in Claude Writing — bttyeo · 2026-11-22