Ollama Plugin: If you have ollama installed you can now use that to server GGUF models instead of llama.cpp
Generate tab: Generate your own datasets using documents, raw text or from scratch. (You can use to train!)
Documents tab: Dedicated page to upload and maintain all documents (for RAG or Generate)
Evals: Downloadable detailed reports for eval tasks, ability to search for tasks
Training: New python completion recipe, several recipe optimizations, added learning rate algorithms to llama_trainer
Bug fixes including:
- Removed dependencies on unzip, detect if curl is missing
- Helpful message when you don’t have an available inference engine
- A number of fixes for issues when using LoRA adapters
- logprobs no longer freezes when max tokens is reduced