VMware/open-llama-7b-v2-open-instruct
VMware/open-llama-7b-v2-open-instruct is a 7 billion parameter instruction-tuned causal language model developed by VMware. Based on the Open LLaMA 7B v2 architecture, this model is optimized for commercial use and demonstrates improved performance on code-related tasks compared to its predecessor. It is fine-tuned on an enhanced instruction dataset, making it suitable for a variety of instruction-following applications.
Loading preview...
Model Overview
VMware/open-llama-7b-v2-open-instruct is a 7 billion parameter instruction-tuned language model, building upon the Open LLaMA 7B v2 base model. A key differentiator for this version is its commercial viability under a CC BY-SA-3.0 license, making it accessible for a wide range of applications. The model shows improved performance on code generation tasks compared to its v1 counterpart, attributed to enhancements in the base model by the openlm-research team and an improved instruction tuning dataset.
Key Capabilities
- Instruction Following: Trained on an enhanced instruction tuning dataset, including Open-instruct-v1 (Mosaic/Dolly-HHRLHF + filtered OASST1) and a subset of COT SUBMIX (from FLAN V2) zero-shot examples.
- Code Generation: Demonstrates better performance in generating code, as highlighted by the README.
- Commercial Use: Licensed under CC BY-SA-3.0, allowing for commercial applications.
Performance Metrics
Evaluated on the Open LLM Leaderboard, the model achieved an average score of 40.34. Notable scores include 70.31 on HellaSwag (10-shot) and 64.33 on Winogrande (5-shot). While showing strengths in some areas, its GSM8K (5-shot) score is 7.43 and MMLU (5-shot) is 35.16.
Usage Note
Users should be aware that the model was trained using the Alpaca prompt template. When instantiating the tokenizer, it is crucial to set use_fast = False to avoid incorrect encoding issues.