Overview

Blossom-v4-qwen1_5-14b is a 14.2 billion parameter conversational language model developed by Azure99. It is built upon the Qwen1.5-14B pre-trained model and has undergone instruction-tuning using a proprietary blend of high-quality English and Chinese datasets, including Blossom Orca, Wizard, Chat, and Math data. This model is specifically optimized for dialogue and multi-turn conversations, demonstrating robust general capabilities and strong contextual understanding.

Key Capabilities

• Conversational AI: Designed for engaging in natural, multi-turn dialogues.
• Context Understanding: Excels at maintaining context across extended conversations.
• Multilingual Support: Trained on high-quality English and Chinese datasets, making it suitable for bilingual applications.
• Instruction Following: Fine-tuned to accurately follow a wide range of instructions.

Training Methodology

Training was conducted in two stages:

1. Stage 1: 1 epoch on a 220K single-turn instruction dataset (100K Wizard, 100K Orca, 20K Math).
2. Stage 2: 3 epochs on a 50K Blossom chat multi-turn dialogue dataset, supplemented with a 2% random sample from Stage 1 data.

Good for

• Building chatbots and virtual assistants.
• Applications requiring strong multi-turn dialogue capabilities.
• Use cases demanding good context retention in conversations.
• Projects targeting both English and Chinese speaking users.