Hugging Face instructlab/granite-7b-lab: An IBM Research Model for Chat Applications
Granite-7b-lab is a 7 billion parameter language model developed by IBM Research, built upon the Granite-7b-base architecture. It is distinguished by its use of the Large-scale Alignment for chatBots (LAB) methodology, a novel synthetic data-based alignment tuning method. This approach leverages Mixtral-8x7B-Instruct as a teacher model to enhance the model's capabilities.
Key Capabilities & Methodology
The LAB methodology is designed to add new knowledge and skills incrementally to an already pre-trained model, effectively mitigating catastrophic forgetting. It comprises three core components:
- Taxonomy-driven data curation: This process uses a tree of seed examples to prompt the teacher model, ensuring diverse and targeted synthetic data generation across various knowledge domains and skills.
- Large-scale synthetic data generator: Unlike traditional self-instruct methods, LAB samples local examples within leaf nodes of the taxonomy, allowing the teacher model to better exploit task distributions.
- Two-phased training with replay buffers: This training regimen includes distinct knowledge tuning and skills tuning phases, utilizing replay buffers to reinforce learning and prevent forgetting.
Performance & Use Cases
While Granite-7b-lab shows a MTBench (Avg) score of 6.69 and MMLU (5-shot) of 51.91, its primary strength lies in its unique training methodology for continuous improvement and adaptation. It is particularly suited for chat applications where incremental learning and the ability to incorporate new, domain-specific knowledge are crucial. The model is primarily English-language and is released under an Apache 2.0 license. Users should note that as a base model, it has not undergone safety alignment and may produce problematic outputs, requiring careful implementation with appropriate safeguards.