Provider: Meta AI
License: Meta Llama 2 Community License (permissive for commercial use with compliance)
Access: Open weights available via Hugging Face and compatible inference frameworks
Architecture: Transformer decoder
Parameters: 7 billion
🔍 Overview
LLaMA 2 7B (Base) is the pretrained foundation model in Meta’s LLaMA 2 family. It is not instruction-tuned, making it suitable for research, fine-tuning, and downstream adaptation in a wide range of domains — from chat models to coding assistants to domain-specific LLMs.
Key characteristics:
- Compact & Powerful: Offers a strong performance-to-size ratio
- Pretraining-Only: Requires tuning or prompting to perform tasks effectively
- Flexible Backbone: Commonly used for QLoRA, PEFT, and instruction fine-tuning workflows
⚙️ Technical Specs
- Architecture: Decoder-only transformer
- Parameters: 7B
- Context Length: 4K tokens
- Tokenizer: SentencePiece (same as LLaMA 1)
- Training Data: 2 trillion tokens of publicly available text data
🚀 Deployment
- Hugging Face Repo: meta-llama/Llama-2-7b
- Compatible Tools: 🤗 Transformers, text-generation-inference, llama.cpp, PEFT, LoRA/QLoRA
- Fine-Tuning Options: Easily adaptable to specific use cases using open-source tuning libraries