跳到正文
RCreddit.com·
暂不在当前实时榜单

Liked Muse, so I cut the 30B model in half by width, distilled it back, and it does 57 of 60 tool tasks its parent does 60 of

AI 摘要

A developer successfully distilled a smaller model, Ornith-1.0-9B, from the Muse-Glimmer-30B parent model. By reducing the parent's width and using distillation, the 14.44B parameter model achieved 57 out of 60 tool tasks, closely matching the parent's 60 tasks. The distillation process involved 6,000 steps and 98.3M tokens over 162 hours, focusing on tool calling fidelity. This smaller model, available under Apache-2.0, demonstrates effective downscaling while retaining significant functionality, though it's not a general replacement for Muse.

为什么是这条

Unlike other distillation efforts that aim for general replacement, this project specifically focused on retaining tool-calling capabilities, achieving 57 of 60 tasks compared to the parent's 60.

时间与来源

时间显示为 UTC

显示时区:UTC

本地时区尚不可用,暂时显示 UTC。

收录当时偏移:UTC+02026年9月28日 07:00 UTC

收录
2026年9月28日 07:00
来源类型
开发者社区

讨论趋势

→ 平稳
最近 24 小时与此前 24 小时的快照均值对比 · 7 天曲线

百分比基于采集到的讨论信号,不代表新增评论数或独立参与人数。曲线仅用于同一话题在不同时段的比较。

本站未收录正文。

前往源站阅读 →
来源·reddit.com