Internally at Modal, the team decided to experiment with open-source large language models to see if they could streamline their communication on Slack. By fine-tuning a model on Slack messages, they created @erik-bot, which quickly became an invaluable asset offering advice on various topics. Due to the overwhelming success of @erik-bot, they decided to release DoppelBot, allowing others to fine-tune a bot on their own Slack messages. The entire process, from scraping messages to inference, runs on Modal’s serverless components. By utilizing Low-Rank Adaptation for fine-tuning, the team was able to optimize the model efficiently. The team also touches on multi-workspace support for Slack apps.
https://modal.com/docs/examples/slack-finetune