About Machine Learning Model Llama-2
Llama-2 is an advanced collection of pretrained and fine-tuned large language models (LLMs), developed and released by Meta AI. With a parameter range from 7 billion to 70 billion, Llama 2 stands as an updated version of its predecessor, Llama 1. It represents a significant step forward in the field of LLMs, particularly for dialogue use cases through its fine-tuned variants, Llama 2-Chat. This family of models not only surpasses the performance of existing open-source chat models on several benchmarks but also matches the level of some closed-source models in terms of helpfulness and safety, as evaluated by human testing. Llama 2's development involved substantial changes from the original Llama, including an expanded pretraining corpus by 40%, a doubled context length, and the adoption of grouped-query attention. The team behind Llama 2 focused on increasing the safety of these models through specific data annotation and tuning techniques, red-teaming, and iterative evaluations.
Model Card for Llama-2
-
Model Details:
- Developers: Meta AI
- Model Dates: January 2023 to July 2023
- Variations: 7B, 13B, and 70B parameter sizes, both pretrained and fine-tuned
- Architecture: Auto-regressive language model using an optimized transformer architecture. Fine-tuned versions employ supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF).
- License: Custom commercial license (ai.meta.com/resources/models-and-libraries/llama-downloads/)
- Comments: Feedback can be provided through the model README or the GitHub repository (github.com/facebookresearch/llama)
-
Intended Use:
- Primary Uses: Commercial and research applications in English; assistant-like chat for tuned models and various natural language generation tasks for pretrained models
- Out-of-Scope Uses: Any use that violates applicable laws or regulations, use in languages other than English, or any use prohibited by the Acceptable Use Policy and Licensing Agreement for Llama 2
-
Training Data:
- Overview: Trained on 2 trillion tokens from publicly available sources, not including Meta user data. Fine-tuning data includes publicly available instruction datasets and over one million new human-annotated examples.
- Data Freshness: Pretraining data cut off in September 2022, with some tuning data up to July 2023
-
Ethical Considerations and Limitations:
- Llama 2 carries inherent risks due to its new technology. Testing has been limited to English and cannot cover all scenarios. The potential outputs of Llama 2 cannot be predicted in advance and may produce inaccurate or objectionable responses. Developers are advised to perform safety testing and tuning tailored to their specific applications of the model.
-
Hardware and Software:
- Utilized custom training libraries, Meta’s Research Super Cluster, and production clusters for pretraining. Fine-tuning, annotation, and evaluation were also performed on third-party cloud compute.
-
Carbon Footprint:
- Pretraining involved 3.3M GPU hours on A100-80GB hardware. Estimated total emissions were 539 tCO2.