I trained a 348M model trained from scratch on 22.7B tokens that does 14 digit arithmetic [P]

A developer trained a 348M parameter language model from scratch on 22.7B tokens. This model, referred to as "350M V3 instruct," was then fine-tuned to perform 14-digit arithmetic. Unlike models that guess answers, this model solves problems by demonstrating the steps, including column addition with carries, borrow chains, and partial-product multiplication. The developer noted that this was their fifth small language model and part of their third series.