Original model link: Pclanglais/ScikitLLM-Model.
For imatrix data generation, kalomaze's groups_merged.txt were used, you can find it here.
Original model README below.
ScikitLLM is an LLM finetuned on writing references and code for the Scikit-Learn documentation.
Features of ScikitLLM includes:
- Support for RAG (three chunks)
- Sources and quotations using a modified version of the wiki syntax ("")
- Code samples and examples based on the code quoted in the chunks.
- Expanded knowledge/familiarity with the Scikit-Learn concepts and documentation.
Training
ScikitLLM is based on Mistral-OpenHermes 7B, a pre-existing finetune version of Mistral 7B. OpenHermes already include many desired capacities for the end use, including instruction tuning, source analysis, and native support for the chatML syntax.
As a fine-tune of a fine-tune, ScikitLLM has been trained with a lower learning rate than is commonly used in fine-tuning projects.