Background: Using real-world user data from a previous farmer assistant chatbot service and additional curated datasets (prioritizing sustainable regenerative organic farming practices,) Gemma 2B and Mistral 7B LLMs were iteratively fine-tuned and tested against eachother as well as basic benchmarking, whereby the Gemma 2B fine-tune emerged victorious. LORA adapters were saved for each model. Following this, the Gemma version was released.

Updates for this model: We then revisited the data, adding four additional months of real-world in-field data from hundreds of users which was then editted by a domain expert in regenerative farming and natural farming (approximately 2,000 instruct examples.) This was combined with a small portion of synthetic datasets and semisynthetic datasets related to regenerative agriculture and natural farming, including some non-english language samples. The results were far superior to our previous releases of Natural Farming Gemma and Mistral fine-tunes. This Natural Farming Mistral 7B fine-tune was the best scoring in our prelim benchmarking, and so it was converted to GGUF and loaded onto Hugging Face Hub as our Natural Farming Mistral V4 in hopes it will help farmers everywhere and inspire future works.

Shout out to roger j (bhugxer) and Steve Raisner for help with currating the dataset.

Testing and further compiling to integrate into on-device app interfaces are ongoing. This project was created by Copyleft Cultivars, a nonprofit, in partnership with Open Nutrient Project and Evergreen State College. This project serves to democratize access to farming knowledge and support the protection of vulnerable plants.

This is V4 beta. It runs locally on Ollama with some expirimental configuring so you can use it off the grid and places where internet is not accessible (ie most farms I've been on.)

This mistral model was trained with Unsloth and Huggingface's TRL library.