| Property | Value |
|---|---|
| License | CC-BY-NC-4.0 |
| Training Data | 106B tokens of scientific text + Alpaca dataset |
| Training Resources | 16 A100 80GB GPUs, 6 hours training |
| Framework | PyTorch/Transformers |
What is galpaca-30b?
GALPACA-30B is an advanced language model that combines the scientific expertise of GALACTICA with the instruction-following capabilities of Alpaca. It's specifically designed for scientific and technical tasks, created by fine-tuning the GALACTICA 30B model on Stanford's Alpaca dataset of 52k instruction-response pairs.
Implementation Details
The model was trained using DeepSpeed ZeRO Stage 3 optimizations with 16-bit mixed-precision training. It utilizes a maximum context window of 384 tokens and was trained with an effective batch size of 1024.
- Built on GALACTICA's foundation of 106 billion tokens of scientific text
- Fine-tuned using the Alpaca dataset for improved instruction following
- Implements the Transformers library architecture
Core Capabilities
- Superior performance on technical and scientific tasks
- Strong programming and mathematical reasoning abilities
- Effective instruction following for scientific queries
- Ability to handle complex mathematical notation and formulas
Frequently Asked Questions
Q: What makes this model unique?
GALPACA-30B stands out for its specialized scientific knowledge combined with instruction-following capabilities. It outperforms LLaMA-based Alpaca models on technical tasks while maintaining scientific accuracy.
Q: What are the recommended use cases?
The model is best suited for scientific research, technical documentation, mathematical problem-solving, and programming tasks. However, it's important to note that it's licensed for non-commercial use only and should not be used in production without proper safeguards.