Faradaylab/ARIA-70B-V2
Faradaylab/ARIA-70B-V2 is a 69 billion parameter Llama 2-based generative text model developed by FARADAY. It is specifically fine-tuned on over 50,000 high-quality French tokens to enhance its performance and general topic understanding in the French language. This model is optimized for dialogue use cases and can handle large files for data extraction through experimental rope scaling to increase context length.
Loading preview...
Faradaylab/ARIA-70B-V2: French-Optimized Llama 2
ARIA-70B-V2 is a 69 billion parameter language model developed by FARADAY, built upon the Llama 2-70B-Chat-HF architecture. Its primary distinction lies in its extensive fine-tuning on a proprietary dataset of over 50,000 high-quality French tokens, aiming to significantly improve its proficiency and general knowledge in the French language. This fine-tuning process involved removing Alpaca-style translated English text from the dataset to ensure native French quality.
Key Capabilities & Features
- French Language Optimization: Specifically trained on a large French dataset to enhance performance in French dialogue and general topics.
- Llama 2 Foundation: Benefits from the robust Llama 2 architecture, optimized for dialogue use cases.
- Experimental Rope Scaling: Features an experimental approach to increase context length from 4,096 to over 6,000 tokens, enabling the model to process larger documents for tasks like data extraction.
- Dialogue Optimized: Inherits Llama 2's fine-tuning for human preferences in helpfulness and safety through SFT and RLHF.
Good For
- Applications requiring high-quality French language generation and understanding.
- Dialogue systems and chatbots operating in French.
- Tasks involving data extraction from large French text files, especially with activated rope scaling.
- Developers seeking a powerful, French-centric LLM based on a proven architecture.