TigerResearch/tigerbot-70b-base-v1 Overview
TigerResearch/tigerbot-70b-base-v1 is a 70 billion parameter foundational large language model developed by TigerResearch. It is designed to serve as a robust base model, providing a strong starting point for developers to build and fine-tune their own specialized LLMs. The model's architecture and training focus on general language understanding and generation, making it adaptable to a wide array of downstream tasks.
Key Capabilities and Performance
This base model demonstrates competitive performance across several standard benchmarks, indicating its general utility:
- Average Score: Achieves an average score of 62.1 on the Open LLM Leaderboard evaluation.
- Reasoning: Scores 62.46 on ARC (25-shot) and 37.76 on GSM8K (5-shot).
- Common Sense: Performs well on HellaSwag (10-shot) with 83.61 and Winogrande (5-shot) with 80.19.
- Knowledge & Understanding: Attains 65.49 on MMLU (5-shot) and 52.76 on TruthfulQA (0-shot).
- Reading Comprehension: Scores 52.45 on DROP (3-shot).
Intended Use Cases
TigerResearch/tigerbot-70b-base-v1 is primarily intended as a foundation model for:
- Further Fine-tuning: Developers can fine-tune this base model for specific domains, tasks, or instruction-following capabilities.
- Research and Development: It provides a powerful backbone for exploring new LLM applications and architectures.
- Custom LLM Creation: Users looking to build their own proprietary LLMs can leverage this model as a starting point, customizing it to meet unique requirements.