Forecasting Granite Final is a fine tuned version of Granite 3.3 2b Instruct.
It was fine tuned on a one gigabyte dataset compiled from five gigabytes of internet scraped data on the topic of event forecasting and problem solving using LlamaFactory over three days on a single RX 7900 XTX gpu.
Using the method of freeze training, the first five layers were unfrozen and trained at a learning rate of 1e-4 with two epochs.
This model was designed to run as the main forecasting LLM for the Metaculs AI Forecasting competition to replace the big, inefficient, non-specialized LLMs provided by Metaclus (or self provided for that matter) as there are very few LLms out there designed specifically for event forecasting. This model's accuracy hasn't been tested, and whether or not the fine tuning improved its forecasting capabilities is not known. An interesting trait of this model is that it "talks" in dot-point explanations, even when explicitly told not to. For the use-case of the model, this shouldn't pose any issues.
TEMPERATURE MUST BE SET TO 0 FOR IT TO FUNCTION. The model devolves into loops when given complex and long tasks, a temperature of 0 or close to 0 seems to fix this.
- Downloads last month
- -