Hugging Face Quyen-v0.1: An Instruction-Tuned Qwen1.5 Model
Quyen-v0.1 is a 4 billion parameter model from the Quyen series, developed by vilm. This series is built upon the Qwen1.5 architecture and includes various sizes, with Quyen-v0.1 representing the mid-range offering. The model has undergone extensive instruction tuning using both Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
Key Training Details
- Base Architecture: Qwen1.5 family.
- Fine-tuning Methods: SFT and DPO.
- Training Datasets: A combination of publicly available and private datasets, including:
- OpenHermes-2.5 by Teknium
- Capyabara by LDJ
- argilla/distilabel-capybara-dpo-7k-binarized by argilla
- orca_dpo_pairs by Intel
- Private data from Ontocord and BEE-spoke-data.
Prompt Template
All Quyen models, including Quyen-v0.1, utilize the ChatML format for prompting. This standardized template ensures consistent interaction and optimal performance with the model. Developers can easily integrate this by applying the chat template via the tokenizer.
Use Cases
Quyen-v0.1 is suitable for a wide range of conversational AI applications, instruction following, and general text generation tasks, benefiting from its robust fine-tuning on diverse preference datasets. Its 4 billion parameters offer a balance between performance and computational efficiency.
Note: Benchmark results for the Quyen series are currently pending and will be released at a later date.