Llama 2
lastmod 2023-08-04created 2023-07-213 edits
Meta releases its next-generation language model Llama 2
Features
- Open-source large language model usable for both research and commercial purposes
- Trained on 2 trillion tokens of data, 40% more than Llama 1
- Twice the context length: 4096
- Models in three sizes: 7B, 13B, 70B
- Outperforms other open-source language models (MPT, Falcon, Llama-1) on many benchmarks, including reasoning, coding, proficiency, and knowledge tests
- Trained on publicly available data sources; the fine-tuned Llama-2-chat also uses publicly available instruction sets and over one million human annotations
- Included in the download
- Model code
- Model Weights
- README (User Guide)
- Responsible Use Guide
- License
- Acceptable Use Policy
- Model Card
Running Llama 2 Uncensored locally
ML engineer Eric Hartford wrote a blog post titled "Uncensored Models" in May 2023, which became popular.
Summarizes the benefits of running uncensored models and how to create them
- How to run uncensored Llama 2 models, and a comparison with censored models
- Uncensored models currently available
- Fine-tuned Llama 2 7B model
- Uses the uncensored/unfiltered Wizard-Vicuna conversation dataset
- Fine-tuned Llama 2 7B model
- Llama 2 13B model fine-tuned on 300,000 instructions
- Characterized by long responses, a lower hallucination rate, and no OpenAI censorship mechanism
- Llama 1 13B model fine-tuned to remove alignment
- Movies, cooking (a recipe for dangerously spicy mayonnaise), religion, medicine (how is Tylenol made?), general information (who would win if Mark Zuckerberg and Elon Musk had a fight?)