QuelLLM
A documentation aggregation platform providing guides for local deployment and selection of large language models, focused on the evaluation and testing of French and Francophone AI models.
Timeline (last 90 days)
Developer nationality, French generation quality, and data processing location are three mutually independent dimensions and cannot be conflated.
Downloading the model weight files itself does not constitute proof that data stays local; one must run in an offline environment and check the complete application chain including the inference engine, document retrieval, extensions, and sync functionality.
QuelLLM suggests testing the 3B and 8B models with the same set of texts, with test dimensions including faithful rewriting, precise extraction (JSON format), and missing-information identification, and separately recording accuracy, format compliance, French naturalness, and response time.
When choosing between 3B and 8B, if resources are limited one should start with 3B, then use the same testing process to try 8B, and should not infer quality level from parameter count alone.
AI developed in France is not necessarily stronger in French, and the developer's place of origin cannot substitute for evaluation based on one's own texts.
All Facts (5)
Developer nationality, French generation quality, and data processing location are three mutually independent dimensions and cannot be conflated.
50%UnverifiedDownloading the model weight files itself does not constitute proof that data stays local; one must run in an offline environment and check the complete application chain including the inference engine, document retrieval, extensions, and sync functionality.
50%UnverifiedQuelLLM suggests testing the 3B and 8B models with the same set of texts, with test dimensions including faithful rewriting, precise extraction (JSON format), and missing-information identification, and separately recording accuracy, format compliance, French naturalness, and response time.
50%UnverifiedWhen choosing between 3B and 8B, if resources are limited one should start with 3B, then use the same testing process to try 8B, and should not infer quality level from parameter count alone.
50%UnverifiedAI developed in France is not necessarily stronger in French, and the developer's place of origin cannot substitute for evaluation based on one's own texts.
50%