Azure99/Blossom-V6.4-27B
Azure99/Blossom-V6.4-27B is a 27 billion parameter open-source conversational large language model developed by Azure99, featuring a 32768-token context length. It is designed for general-purpose conversational tasks, leveraging a unique data synthesis workflow that employs multiple advanced LLMs for response generation and verification. This model aims to provide a powerful and cost-effective locally accessible solution, maintaining multimodal capabilities from its base models.
Loading preview...
Blossom-V6.4-27B Overview
Blossom-V6.4-27B is an open-source, 27 billion parameter conversational large language model developed by Azure99. It builds upon the V6.3 training recipe, utilizing the same training data with additional multimodal samples to preserve its inherent multimodal capabilities. The model is designed to be a powerful, cost-effective, and locally accessible general-purpose solution.
Key Capabilities & Differentiators
- Advanced Data Synthesis: Employs a sophisticated, cost-effective data synthesis workflow using Deepseek-V3.1, Gemini 2.5 Flash, and Qwen3-235B-A22B-Instruct-2507. This process involves multi-model verification for objective scenarios (e.g., mathematics) and cross-evaluation for subjective scenarios to generate high-quality training data.
- Reproducible Post-Training Data: Focuses on providing reproducible post-training data, enhancing transparency and reliability.
- Multimodal Preservation: Incorporates multimodal samples to ensure the retention of multimodal capabilities from its base models.
- Robust Filtering: Utilizes rule-based filtering, including N-Gram filtering for repetition removal and discarding toxic content, to refine the training dataset.
Training Data
The model's training data largely follows the V6.3 series, with specific stages available on Hugging Face: Blossom-V6.3-SFT-Stage1 (1 epoch) and Blossom-V6.3-SFT-Stage2 (3 epoch). The data synthesis is managed by the 🌸BlossomData framework.