Atlys reposted this
We fine-tuned a 2B vision-language model to parse bank statements at production scale. The full write-up from Shubham Tiwari is on our engineering blog. Fine-tuning is no longer the hard part. You have the GPU, the dataset, the model. The training script is a few hundred lines. The hard part is everything that happens after. The post covers how we worked through GPU bottlenecks when bank statements pushed output tokens past our context window and resource constraints, tuned the VLM-specific parameters that defaults get wrong, fine-tuned training prompts to fix specific field regressions, and a few other things that mattered more than hyperparameters. Why does this matter? Frontier APIs handle document AI well, until you're processing tens of thousands of pages a day. At that point you're paying thousands a month, inheriting rate limits, and can't fix specific failure modes. A purpose-built 2B model on a single GPU changes the economics. Really proud of the work Shubham and the team put into this. https://lnkd.in/g3qkuQvd