Tencent/digitalhuman
May 2025 – May 2025
1 Month active
Languages Used
Python
Technical Skills
DebuggingDeep LearningMachine LearningModel DevelopmentModel ImplementationModel Training
mingxiaoli
PROFILE
Mingxiaoli
Mingxiao Li enhanced the Tencent/digitalhuman repository by extending the Llava model’s forward method to support labels and text embeddings, enabling more flexible model training and evaluation. Using Python and deep learning frameworks, Mingxiao introduced a dynamic loss-control mechanism that allows switching between loss calculation strategies through a new method attribute and setter function. To improve integration reliability, Mingxiao fixed import paths for Llama components, ensuring correct module usage, and addressed loss tensor dtype mismatches to stabilize training in the Deepseed environment. The work demonstrated strong debugging and model development skills, resulting in a more robust and adaptable machine learning pipeline.
Overall Statistics
Feature vs Bugs
33%Features
Repository Contributions
4Total
Bugs
2
Commits
4
Features
1
Lines of code
30
Activity Months1
Your Network
155 people
Work History
May 2025
4 Commits • 1 Features
May 1, 2025Monthly work summary for 2025-05 for Tencent/digitalhuman focused on delivering core model enhancements, stabilizing training, and improving integration reliability. Key features delivered include extending the Llava model forward to support labels and text embeddings, and introducing a dynamic loss-control mechanism to switch loss strategies. Major bug fixes addressed import path correctness for Llama components and ensured loss dtype integrity during training, enhancing stability in the Deepseed environment.
4 Commits • 1 Features
May 1, 2025Monthly work summary for 2025-05 for Tencent/digitalhuman focused on delivering core model enhancements, stabilizing training, and improving integration reliability. Key features delivered include extending the Llava model forward to support labels and text embeddings, and introducing a dynamic loss-control mechanism to switch loss strategies. Major bug fixes addressed import path correctness for Llama components and ensured loss dtype integrity during training, enhancing stability in the Deepseed environment.
May 2025
Activity
Loading activity data...
Quality Metrics
Correctness85.0%
Maintainability90.0%
Architecture80.0%
Performance70.0%
AI Usage25.0%
Skills & Technologies
Programming Languages
Python
Technical Skills
DebuggingDeep LearningMachine LearningModel DevelopmentModel ImplementationModel TrainingNatural Language Processing
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline