Core Concepts
Vietnamese PhoGPT models offer state-of-the-art performance and open-source availability for language generation tasks.
Abstract
1. Abstract:
- Open-sourcing a 4B-parameter generative model series for Vietnamese.
- Includes PhoGPT-4B and its chat variant, PhoGPT-4B-Chat.
2. Introduction:
- Success of large language models in English.
- Release of PhoGPT models for Vietnamese.
- Training details and performance comparison.
3. PhoGPT:
3.1 PHOGPT-4B: MODEL ARCHITECTURE AND PRE-TRAINING
- Transformer decoder-based model with specific configurations.
- Pre-training on a diverse corpus of Vietnamese texts.
3.2 PHOGPT-4B-CHAT: SUPERVISED FINE-TUNING
- Fine-tuning process using instructional prompts and conversations from various sources.
4. Evaluation:
- Comparison with closed-source and open-source models.
- Accuracy results favoring PhoGPT models, especially in Vietnam-specific questions.
5. Conclusion:
- Introduction of open-source LLMs for Vietnamese.
Limitations:
PhoGPT limitations include reasoning, coding tasks, generating harmful content, or biased responses.
Stats
PhoGPTのモデルサイズは約3.7Bパラメータです。
PhoGPTは、102Bトークンのベトナム語コーパスで2エポックのために事前学習されました。