vilm/Quyen-Plus-v0.1
Quyen-Plus-v0.1 is a 7.7 billion parameter large language model developed by vilm, based on the Qwen1.5 family. This instruction-tuned model, trained with SFT and DPO on a diverse dataset including OpenHermes-2.5 and Capyabara, is designed for general-purpose conversational AI. It achieves an average score of 63.27 on the Open LLM Leaderboard, demonstrating capabilities across reasoning, common sense, and language understanding tasks.
Loading preview...
Quyen-Plus-v0.1 Overview
Quyen-Plus-v0.1 is a 7.7 billion parameter instruction-tuned large language model, part of the Quyen series developed by vilm. It is built upon the Qwen1.5 architecture and has been fine-tuned using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
Key Capabilities & Training
The model was trained on a comprehensive dataset, including well-known resources like OpenHermes-2.5 by Teknium, Capyabara by LDJ, argilla/distilabel-capybara-dpo-7k-binarized by argilla, and orca_dpo_pairs by Intel, alongside private data from Ontocord and BEE-spoke-data. This diverse training regimen aims to enhance its conversational and reasoning abilities.
Performance Highlights
Evaluated on the Open LLM Leaderboard, Quyen-Plus-v0.1 achieved an average score of 63.27. Specific benchmark results include:
- AI2 Reasoning Challenge (25-Shot): 55.72
- HellaSwag (10-Shot): 78.52
- MMLU (5-Shot): 60.45
- TruthfulQA (0-shot): 53.60
- Winogrande (5-shot): 71.27
- GSM8k (5-shot): 60.05
These scores indicate its proficiency in various tasks, from common sense reasoning to mathematical problem-solving. The model utilizes the ChatML prompt template for interaction, supporting a clear and structured conversational format.
When to Use This Model
Quyen-Plus-v0.1 is suitable for developers seeking a capable 7B-class model for general conversational AI applications, instruction following, and tasks requiring reasoning and language understanding. Its balanced performance across multiple benchmarks makes it a versatile choice for a range of use cases.