| Item | Specifications |
|---|---|
| Model depth (number of layers) | From 32 to 100 |
| Embedding size | From 1024 to 12288 |
| Item | Specifications |
|---|---|
| VMLU | Score of 75 or higher |
| Item | Specifications |
|---|---|
| Context and flow understanding | The model can maintain context in conversation |
| Coherent and logical text generation | Can generate text based on a given format |
| Multilingual processing (if any) | Can process Vietnamese and English |
| Item | Specifications |
|---|---|
| Logical reasoning | Solve simple problems and puzzles |
| Understand and generate code | Can generate simple code |
| Item | Specifications |
|---|---|
| Sensitive content moderation mechanism | Yes |
| Avoid answering content outside the data scope | Yes |
| Item | Specifications |
|---|---|
| Average response time | Depends on length of input and output. Average response time for first token with 1500 tokens input is under 1s |
| Scalability | Capability to scale across multiple GPU clusters |
| Integration with other systems |
- Provides integration capability via API - Capability to integrate RAG for specialized knowledge |
| Item | Specifications |
|---|---|
| Fine-tuning with user data | Allows model customization and adding more training text data |
| Item | Specifications |
|---|---|
| Fast training data creation | Support building training data by importing Excel files |
| Multi-platform chat and internal communication channel integration | Facebook, Zalo, Viber, Telegram, Website, and internal communication channels |
| Model update and development roadmap | Regularly update to new versions. Customers get to use our latest features and models |
| User documentation | PDF documentation available |