Google translategemma 4B Translation Models
2 months ago
- #Hugging-Face
- #Gemma
- #machine-translation
- Accessing Gemma on Hugging Face requires accepting Google's usage license.
- TranslateGemma is a lightweight, state-of-the-art open translation model by Google, supporting 55 languages.
- Inputs include text strings and images (896x896 resolution), with a total input context of 2K tokens.
- Outputs are translations into the target language.
- Usage involves specific chat templates for text or image inputs, requiring source and target language codes.
- Models can be used via Hugging Face pipelines or direct initialization with transformers.
- Training utilized 4.3B tokens for SFT and 10.2M tokens for reinforcement learning, leveraging TPU hardware.
- Evaluation metrics show improvements in safety and performance over previous models.
- Limitations include potential biases, factual inaccuracies, and challenges with complex language nuances.
- Ethical considerations focus on bias mitigation, content safety, and responsible usage guidelines.