| Property | Value |
|---|---|
| Base Model | Qwen2.5-32B-Instruct |
| Parameters | 32B |
| Training Hardware | 8x H100-80G GPUs |
| Transformers Version | 4.46.3 |
| Model URL | huggingface.co/Saxo/Linkbricks-Horizon-AI-Avengers-V1-32B |
What is Linkbricks-Horizon-AI-Avengers-V1-32B?
This is a sophisticated multilingual language model developed by Saxo (Yunsung Ji) at Linkbricks, achieving Rank-2 in the Open LLM Leaderboard for models under 32B parameters. The model underwent a comprehensive fine-tuning process involving CPT, SFT, DPO, and ORPO training on approximately 35% of its parameters.
Implementation Details
The model was trained using an impressive dataset of 80 million multi-lingual news articles and wiki corpus, specifically enhanced for cross-language understanding between Japanese, Korean, Chinese, and English. Technical implementation includes Deepspeed Stage=3, rslora, and BAdam Layer Mode for optimal performance.
- Maintains original tokenizer without word expansion
- Implements Function Call and Tool Calling support
- Utilizes advanced training techniques including CPT->SFT->DPO->ORPO pipeline
Core Capabilities
- High-dimensional analysis of customer reviews and social media posts
- Advanced coding and writing capabilities
- Enhanced mathematical and logical reasoning
- Cross-lingual processing and understanding
- Complex problem-solving in multiple languages
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its sophisticated multi-lingual capabilities combined with enhanced logical reasoning, particularly in East Asian languages. The extensive training on 80M diverse corpus and specialized fine-tuning process makes it particularly effective for cross-lingual tasks.
Q: What are the recommended use cases?
This model excels in multilingual content analysis, customer review processing, social media analysis, coding tasks, and complex mathematical/logical problem-solving. It's particularly suitable for applications requiring deep understanding across Japanese, Korean, Chinese, and English languages.