deepseek-v4.1POST
去审查版 V4.1-Flash-1017 的 heretic 消融实测进行中
作者对 DeepSeek V4.1-Flash-1017 做 heretic 消融得到去审查版,称 MMLU 与 HumanEval 无可测差异,正跑完整基准看工具调用能力,明天再用 BFCL 对比基线。
Kev
@Kevrsub5 天前收录原推发布于 2026.10.25
Kev @Kevrsub原推发布于 2026.10.25
Doing a shake out on an uncensored version of Deepseek V4.1-Flash-1017 with a heretic ablation applied. No measurable difference in MMLU or HumanEval and running a full bench on it now to see if I lost any tooling ability. I'll also run Berkely's BFCL evaluation on it tomorrow and compare it to the baseline. I wanted my own MLX oQ4e version.
x.com/kevrsub/status/2114178727860924618采集于 2026.10.25 12:01