NostraEmpire/mirror-granite-3.1-8b-instruct
Granite-3.1-8B-Instruct is an 8 billion parameter long-context instruct model developed by IBM, fine-tuned from Granite-3.1-8B-Base. It is optimized for solving long-context problems and general instruction following, supporting a context length of 32768 tokens. The model is designed for building AI assistants across various domains, including business applications, and supports 12 languages. It achieves an average score of 71.31 on the HuggingFace Open LLM Leaderboard V1 benchmarks.
Loading preview...
Model Overview
Granite-3.1-8B-Instruct is an 8 billion parameter instruct model developed by IBM, fine-tuned from its base version. It leverages a combination of open-source instruction datasets, internally collected synthetic data, and a structured chat format. The model's development incorporates supervised fine-tuning, reinforcement learning for alignment, and model merging techniques.
Key Capabilities
- Long-Context Tasks: Specifically tailored for long document summarization, long document QA, and other long-context problems.
- Multilingual Support: Supports English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
- General Instruction Following: Designed to respond to a wide range of instructions for building AI assistants.
- Diverse NLP Tasks: Proficient in summarization, text classification, text extraction, question-answering, Retrieval Augmented Generation (RAG), code-related tasks, and function-calling.
Performance Highlights
On the HuggingFace Open LLM Leaderboard V1, Granite-3.1-8B-Instruct achieved an average score of 71.31, with notable scores including 65.34 on MMLU and 73.84 on GSM8K. For more advanced benchmarks (V2), it scored 72.08 on IFEval and 34.09 on BBH.
Architecture
Based on a decoder-only dense transformer architecture, it features GQA, RoPE, MLP with SwiGLU, RMSNorm, and shared input/output embeddings. It was trained on IBM's Blue Vela supercomputing cluster using NVIDIA H100 GPUs.