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

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