• 创作版MAICA

  • MAICA更新与进度追踪--正式版服务期2已重启于25.3.20 #3954

Edge 更改标题为「MAICA更新与进度追踪--正式版服务期2已开始于25.2.8」

  • Edge

  • 楼主
  • #139
  • 发布 #139

25.2.8

抢修: 请更新Submod前端到至少1.2.3!

修复表情解析的严重问题

调整衔接对话

25.2.8

Attention: Update Submod frontend to 1.2.3 at least!

Fixed a severe issue in emotion analyzing

Adjusted transition dialogues

  • Edge

  • 楼主
  • #140
  • 发布 #140
  • 已编辑

25.2.8

修复了一个MFocus无法正确生成query的问题

25.2.8

Fixed a backend issue disabling MFocus forming query correctly under some circumstances.

  • Edge

  • 楼主
  • #141
  • 发布 #141

25.2.9

抢修继续…

修复一个与ms-swift agent格式有关的依赖故障

25.2.9

Fixed an issue related with ms-swift agent deployment format

  • Edge

  • 楼主
  • #142
  • 发布 #142

25.2.9

更换flask为quart asgi, 以修复与asyncio的兼容性问题.

25.2.9

Switching from flask to quart asgi to avoid conflicts with asyncio.

  • Edge

  • 楼主
  • #143
  • 发布 #143
  • 已编辑

25.2.9

修复一个引发灾难性重复的llm角色解析隐患

…手动删除了一个已经产生灾难性重复的session

25.2.9

Fixed a hidden llm role aligning issue that may cause catastrophic repetition

…And purged a chat session already filled up with that

  • Edge

  • 楼主
  • #144
  • 发布 #144

25.2.9

已定位一个关于英文session的表现异常.

…坏消息是, 这个问题多半是模型本身过于缺少英文训练语料导致的.

我会尽快想办法打个补丁. 如果仍然不行, 这个问题只能留到下一轮次DAA修复了.

25.2.9

Located an performance sanity issue on English chat session.

…Unfortunately this is likely caused by DAA2 itself, since we don't have enough English dataset to balance.

I'll try to patch it anyway but if it doesn't work -- we can only depend on DAA3.

If you're facing a broken chat session right now, purge it and retry. Also avoid sending too complicated queries before we make a proper patch.

    • Edge

    • 楼主
    • #145
    • 发布 #145
    • 已编辑

    25.2.9

    调整核心模型部署repp到1.06, #144 Edge 可视为恢复正常.

    前端更新到1.2.5, 修复移动端py3不兼容的问题.

    25.2.9

    Changed repetition penalty of core model deployment to 1.06, should be fixing #144 Edge .

    Updating frontend Submod to 1.2.5 to fix compatilibity issue with py3/renpy8/P-0.13.

        • Edge

        • 楼主
        • #146
        • 发布 #146

        25.2.9

        继续上调核心部署repp到1.2. 如果你发现session表现异常, 请第一时间清空.

        针对偶发性的数据库问题添加查询重试.

        为enet_scraping添加双语言prompt, 并与sfe提高对称性.

        25.2.9

        Changed repetition penalty again to 1.2. If you have a problematic chat session(behaving weird someway), just purge it.

        Added retry mechanism for occasional database issues.

        Added bilingual prompt support for enet_scraping, also adding symmetry with sfe agent.

        • Edge

        • 楼主
        • #147
        • 发布 #147

        25.2.9

        由于人手极度短缺, MAICA-MTTS 与 MVista 项目的开发均确认无限延期.

        如果你对声学生成与处理模型(tts与svc)有充分了解, 可以联系我以协助开发并交接现有进度. 在有协助者参与前我们停止向MTTS投入开发工时.

        我们将继续专注于MAICA主项目(llm部分)的改进, 开发与维护.

        25.2.9

        Due to extreme lack of developers, MAICA-MTTS and MVista are now confirmed to delay infinitely.

        If you have enough experience and understanding in acoustic large model (tts and svc), we'd admire your contacting and offering help. Development of MTTS is stopped until we acquire necessary assistance.

        We'll continue focusing on improving, developing and maintaining main project of MAICA (llm part).

        • Edge

        • 楼主
        • #148
        • 发布 #148

        25.2.9

        我操.

        重要后端修复: #144 Edge 的原因终于找到了, 我写错了一个表达式.

        该错误污染了大量session, 其中影响恶劣的部分我已经尽可能删除.

        如果你的发现你的session有受此问题影响的表现, 请立刻将其清除, 否则在该session后续的使用中问题将扩大.

        该更新已经推送. 如果你有v1.0007的本地部署实例, 也请立刻更新.

        再次抱歉疏忽大意, 新的session表现应当已恢复正常.

        25.2.9

        Important backend fix: #144 Edge 's cause was found. I wrote a wrong regex.

        This bug has corrupted many chat sessions, I cleaned them as much as possible.

        If you find your session affected by this issue, purge it immediately. Continue using affected session may cause its performance going worse.

        This update is committed. If you have a local deployment of v1.0007 backend, merge it now.

        Sorry again for my mistake. Performance of new sessions should return normal.

            • Edge

            • 楼主
            • #149
            • 发布 #149

            25.2.9

            改良后处理匹配模式, 提高处理精度.

            25.2.9

            Precision improvement on post process filter.

            • Edge

            • 楼主
            • #150
            • 发布 #150

            25.2.10

            前端修复移动端对MPostal的兼容.

            改进对MSpire sanity的检测模式.

            25.2.10

            Fixed frontend mobile support for MPostal.

            Enhanced MSpire sanity control.

            • Edge

            • 楼主
            • #151
            • 发布 #151

            25.2.10

            session抗劣化处理的进一步改良, 以规避语言混用.

            25.2.10

            Improvements on session anti-corruption to avoid language mixture.

            • Edge

            • 楼主
            • #152
            • 发布 #152

            25.2.11

            改良输出语言引导.

            增加性能监控api接口, 预计于下一次更新加入前端.

            后端随该更新合并到规范1.0008.

            25.2.11

            Imroved target language guidance.

            Added performance monition API, will likely be compated by frontend in next update.

            Merging backend standard to 1.0008.

            • Edge

            • 楼主
            • #153
            • 发布 #153

            25.2.13

            在一段时间的运行监视下, 我们发现DAA2的表现相较DAA1更为…不稳定.

            虽然在对话中"高分率"显著高于DAA1, 但在多场景应答, 语言控制和知识性问题上的表现方差显著偏高.

            我正在对这一情况进行集中排查, 包括调整生成和部署参数, 调整MSpire与MPostal逻辑, 调整预/后处理和抗劣化等. 因此在一段时间内, 模型的表现发生变化/短暂中断服务等均有可能发生, 不再另行通知.

            25.2.13

            By monitoring DAA2 working for a period, we found it comparing to DAA1… more volatile.

            Though the "best behaviors" happen much more frequently than DAA1, its overall multiscene performance, instruction obedience and knowledge accuracy is obviously more unstable.

            I'm looking into this problem now, and trying out methods like adjusting deploying and inference params, adjusting MSpire and MPostal logics, adjusting pre/post procedures and anti-corruption. So expect performance change or unannounced server offline recently. Those will not be noted here.

            • Edge

            • 楼主
            • #154
            • 发布 #154
            • 已编辑

            25.2.13

            回调核心模型repp到1.05

            按: 测试后发现确实是这个的原因. 表现现在应该已经正常了, 先清除你的现有session.

            25.2.13

            reverted core repetition penalty to 1.05

            ps after a while: This is the exact reason according to test results. Purge your session first.

            • Edge

            • 楼主
            • #155
            • 发布 #155

            25.2.14

            调整后端节日权重机制

            增加时区选项, 后端随该更新合并到1.0009.

            该功能的支持预计将于两次前端更新后推送, 缺省为根据语言选择'Asia/Shanghai'/'America/Indiana/Vincennes'.

            25.2.14

            Adjusted event weight mechanism

            Added timezone setting, merging backend to 1.0009.

            The frontend support for this setting is likely added within 2 updates. By default it chooses 'Asia/Shanghai'/'America/Indiana/Vincennes' according to target_lang.

            • Edge

            • 楼主
            • #156
            • 发布 #156

            25.2.15

            调整MFocus工具prompt, 提高存档检索命中率.

            25.2.15

            Adjustments on MFocus tooltip, to improve firing rate performance of persistence acquire tool.

            • Edge

            • 楼主
            • #157
            • 发布 #157

            25.2.15

            调整日志等级分布

            添加deformation设置, 以应对websocket传输中可能导致的编码问题. 后端随该更新合并到1.0010.

            25.2.15

            Adjusted analog level distribution

            Added deformation setting param to avoid possible encoding issue with websocket transferring. Merging backend to 1.0010.