“2026-02-02”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（7位用户的8个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*
+
* English version for Middle School AI Textbook done.
 +
* AI textbook for old people
 +
* Initial draft for 2nd MD paper.
 
||
 
||
 
*
 
*
第17行： 第19行：
 
|Lantian Li
 
|Lantian Li
 
||
 
||
−
*
+
* MoE daily work (busy week)
 +
* FYT project delivery
 
||
 
||
−
*
+
* Go on the final review of my MLA book (6/10)
 
||
 
||
 
*
 
*
第28行： 第31行：
 
|Wenqiang Du
 
|Wenqiang Du
 
||
 
||
−
*
+
* Collaboration project acceptance (Tencent)
 +
* Year end summary(AIbabel)
 
||
 
||
 
*
 
*
第39行： 第43行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Train MD model with Chinese pretrained front end with cross attention architecture. (Not test yet)
 +
* Trying to figure out actual meaning of existed annotation data with help from xianyu [https://gcndy0j2mmt7.feishu.cn/wiki/LOZKwIQiTiQZdKkJlzEcuOl6n6c]
 
||
 
||
 
*
 
*
第50行： 第55行：
 
|Ying Shi
 
|Ying Shi
 
||
 
||
−
*
+
* Thesis
 
||
 
||
 
*  
 
*  
第111行： 第116行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* GPU Util: [https://z1et6d3xtb.feishu.cn/wiki/XX4NwX3tJiBDcgkMi0hcFUtInHh?from=from_copylink]
 +
* LLM: Finish complex topo code writing, and build swarm with copying node inside it (experiment still running)
 
||
 
||
 
*
 
*
第122行： 第128行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*
+
* LLM: build swarm for  MMLU exp with different LLM base model in FinalDecision: exchange qwen2.5 to llama3-8b
 
||
 
||
 
*
 
*
第165行： 第171行：
 
|Bochao Hu
 
|Bochao Hu
 
||
 
||
−
*
+
* stage2 P2S: finetune with NLLB model using syn data, result is nearly same as before
 +
* read papers
 
||
 
||
 
*
 
*

2026年2月2日 (一) 11:04的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • English version for Middle School AI Textbook done.
  • AI textbook for old people
  • Initial draft for 2nd MD paper.
Lantian Li
  • MoE daily work (busy week)
  • FYT project delivery
  • Go on the final review of my MLA book (6/10)
Wenqiang Du
  • Collaboration project acceptance (Tencent)
  • Year end summary(AIbabel)
Yang Wei
  • Train MD model with Chinese pretrained front end with cross attention architecture. (Not test yet)
  • Trying to figure out actual meaning of existed annotation data with help from xianyu [1]
Ying Shi
  • Thesis
Yue Gu
  • write my Phd thesis and collect the picture
Lily
  • Organized course materials production （小初高分册+ Middle school English version）
  • Participated in reviewing the AI handbooks
  • AIGE 2026 meeting-related tasks
Pengqi Li
  • Currently teaching at the AI Winter Camp. I will be fully back to my research in Friday.
Junming Yuan
  • checking EN middle school AI handbook(M5-M7)
  • learn-not-to-listen MT-HuBERT:
    • setting a threshold can improve performance on clean-speech tasks(ASR, MT-HuBERT:9.22 --> LNTL-MT-HuBERT: 8.43)
    • there is some performance degradation on mixed-speech tasks(SD, MT-HuBERT:3.93 --> LNTL-MT-HuBERT: 4.31).
  • draft of the journal paper(almost done)
Yu Zhang
  • GPU Util: [2]
  • LLM: Finish complex topo code writing, and build swarm with copying node inside it (experiment still running)
Junhui Chen
  • LLM: build swarm for MMLU exp with different LLM base model in FinalDecision: exchange qwen2.5 to llama3-8b
Jiaying Wang
  • spk order experiment finished, 4mix result are under test.
  • loudness & content order verification done.
  • spk order verification code preparing(2/3).
Xiaoxue Luo
  • 2-5mix multi_head separation model for Huawei project
    • Modifications：
      • remove data that the mixed audio equals the target audio
      • add sing data to the speech category
      • add MSE_loss to the original SISDR_loss to control the energy amplitude of output audio
    • Result: compared to the previous model, the performance is even worse
    • Reasons：
      • add sing data to the speech branch is not feasible
      • there are some issues with the istft function in the original paper code
Bochao Hu
  • stage2 P2S: finetune with NLLB model using syn data, result is nearly same as before
  • read papers
Hongcheng Zhang
  • prepare for weekly report
  • read some new paper
  • enrich my train data
Weiman Sun
  • relevant papers
  • hook the first token after the prompt to observe the model's understanding of the task at different depths.
取自“http://index.cslt.org/mediawiki/index.php?title=2026-02-02&oldid=43932”