让变化经过筛选，
再抵达你。

作者对 Agents-A1-4B 的实验报告称，单轮方案以主体权重 2.655 比特存储，在三项基准上平均保留约 93.7% 的原模型分数。该比特数不包含全部参数，推理仍先还原为 BF16；因此不能据此推断显存占用或速度。结果来自作者自测，原始单次比较没有置信区间。

The author's Agents-A1-4B experiment reports roughly 93.7% mean score retention across three benchmarks with body weights stored at 2.655 bits after one round. That rate excludes some parameters, and inference first restores BF16 weights; it does not establish runtime memory or speed. These are author-run evaluations, with no confidence intervals for the original single-run comparison.

YuE2 模型卡标注 9 月 12 日更新，比较 17 种设置，使用 192 条 WildSongBench 提示，并公开指标与协议。作者报告的领先结果采用 best-of-8，即多个候选中选优，而非单次生成。它是团队自己的基准结果，不代表任意曲风、每次采样或所有听众的偏好。

YuE2's model card dates a comparison update to September 12, covering 17 settings on 192 WildSongBench prompts and linking metrics and protocols. The authors' leading result uses best-of-8 selection, not a single generation. This is a team-reported benchmark, not a guarantee across musical styles, individual samples, or listener preferences.

作者在论坛介绍一个含 38 份可运行 notebook 的学习仓库，覆盖分词、手写训练循环、LoRA 和 GRPO，并包含按文章划分训练与测试数据的问答示例。仓库与说明已公开；这是个人学习项目，不是 Hugging Face 官方课程发布，也不代表所有示例已由本站执行验证。

A forum author introduced a learning repository with 38 runnable notebooks on tokenization, manual training loops, LoRA, and GRPO, including a question-answering example with article-level train/test separation. The repository and explanation are public. This is a personal learning project, not an official Hugging Face course release, and we have not executed every example.