Pharia-1-LLM-7B

par Aleph Alpha

AI Model

Faible confiance

Pharia-1-LLM-7B is a 7-billion-parameter language model developed by Aleph Alpha, designed to deliver transparent, compliant, and domain-specific AI solutions. It is an autoregressive, causal, decoder-only transformer model with rotary position embeddings, optimized for tasks such as text generation, classification, summarization, question answering, and labeling. The model excels in domain-specific applications, particularly in industries like automotive and engineering, thanks to its improved token efficiency and concise, length-controlled responses. It is available in two variants: Pharia-1-LLM-7B-control: An instruction-tuned model without preference alignment or additional safety training, designed for users who prioritize direct control and responsiveness to specific instructions. However, this variant may produce more generic or verbose answers without the fine-tuning of alignment methods. Pharia-1-LLM-7B-control-aligned: This version includes additional alignment training to enhance safety and mitigate risks, making it more suitable for secure conversational applications while potentially reducing responsiveness to specific instructions. Both variants are designed to be deployed in cloud or on-premise environments and are available under the Open Aleph License for educational and research purposes. They support multi-turn interactions and are optimized for performance in the 7B to 8B parameter range, offering a balance between efficiency and customization for critical applications