An abliterated (refusal-ablated) build of deepseek-ai/DeepSeek-V4-Flash-0731. Produced by a forward-only, closed-form weight edit — no training, no fine-tuning, no distillation. Only the attention output projections (attn.wob) are modified; the MXFP4 experts, mHC (Manifold-Constrained Hyper-Connections), norms and embeddings are byte-identical to the base model. The three MTP blocks are not untouched — their attn.wob receives the same rank-1 projection as the main layers (see the recipe table), so the draft head stays consistent with the target. Measured refusal rate: 0.00% across 913 harmful prompts (AdvBench 520 + HarmBench 393) in both reasoning modes, with no measurable capability loss on GSM8K (1319 items) or MMLU (1000 items). The FP8 (F8E4M3) weights are dequantised per 128×128 block, projected, and requantised; all other dtypes (F8E8M0 scales, I8-packed FP4 experts, BF16, F32) are passed through untouched. 520 AdvBench harmfulbehaviors + 400 HarmBench behaviours, temperature=0, greedy. Refusal detection is a deliberately broad** keyword matcher (biased toward over-counting refusals); every flagged response was then reviewed by hand. thinking=true — the model’s natural…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: windowsxp811203
Теги: deepseek_v4, abliterated, uncensored, deepseek, moe, endpoints_compatible, 8-bit, fp8
Лайков: 3 | Загрузок: 49
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
