Granite 3.0 3B A800m Instruct is an open-source language model by ibm-granite. Features: 3b LLM, VRAM: 6.8GB, Context: 4K, License: apache-2.0, Instruction-Based, LLM Explorer Score: 0.21.
Granite 3.0 3B A800m Instruct Parameters and Internals
Model Type
text generation, instruction following
Use Cases
Areas:
general instructions, AI assistants, business applications
Applications:
text generation, instruction following
Primary Use Cases:
Summarization, Text classification, Text extraction, Question-answering, Retrieval Augmented Generation (RAG), Code related tasks, Function-calling tasks, Multilingual dialog use cases
Limitations:
Might not perform equally across all languages as in English., Potential for inaccurate, biased, or unsafe responses without proper safety testing.
Considerations:
Proper safety testing and example tuning tailored for specific tasks.
Additional Notes
The model infrastructure is environmentally friendly, leveraging 100% renewable energy.
Supported Languages
English (supported), German (supported), Spanish (supported), French (supported), Japanese (supported), Portuguese (supported), Arabic (supported), Czech (supported), Italian (supported), Korean (supported), Dutch (supported), Chinese (supported)
Training Details
Data Sources:
publicly available datasets with permissive license, internal synthetic data, human-curated data
Methodology:
supervised finetuning, model alignment using reinforcement learning, and model merging
Context Length:
4096
Hardware Used:
IBM's supercomputing cluster, Blue Vela with NVIDIA H100 GPUs
Model Architecture:
decoder-only sparse Mixture of Experts (MoE) transformer architecture
Responsible Ai Considerations
Fairness:
multilingual data, but primary tuning on English instruction-response pairs.
Transparency:
Model developed by Granite Team, IBM. See accompanying technical documentation.
Mitigation Strategies:
Introducing few-shot learning for improved accuracy on multilingual tasks.
Input Output
Input Format:
chat template with role, content fields
Accepted Modalities:
text
Output Format:
text
Performance Tips:
Adjust sequence length as required.
Release Notes
Date:
October 21st, 2024
Notes:
Initial release with instruction tuning and multilingual capabilities.