SFT
Beyond Market Intelligence keeps SFT in one place: 3 stories so far. The section currently leads with “Building a Polish legal AI model starts with one clear choice”, “Finding where AI's telltale writing style is installed during training”, and “Explore language from the past with a focused AI built on vintage text.”. Training a Polish legal AI from a base model is a genuinely ambitious project, and the hardest part isn't the code, it's knowing what to do with your data. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every SFT story on Beyond Market Intelligence, newest first.
Building a Polish legal AI model starts with one clear choice
Training a Polish legal AI from a base model is a genuinely ambitious project, and the hardest part isn't the code, it's knowing what to do with your data. You've already done solid CPT work on legislation and court decisions. Now you're stuck between SFT and RL, unsure how to teach reasoning without trusting a frontier model to get Polish law right. That's the right problem to have. Skip more CPT. Build a clean SFT set from your highest-confidence Q&A first, then layer in tool-use.
Finding where AI's telltale writing style is installed during training
The telltale assistant style isn't scattered randomly through a model: it's a specific direction in the weights, installed at a specific stage. Qwen3-4B-Instruct shows a clear axis separating exaggerated from plain writing, and that axis is already present in the base model. The real surprise is that SFT introduces the rhetorical structures, but DPO doubles the slop score through buzzwords and formatting. Directional ablation actually increases slop by 89%, which tells us clean text projects negatively on the axis.
Explore language from the past with a focused AI built on vintage text.
Three months and $807 bought Unbounded Labs a vintage LLM named Bart, trained from scratch on 20.1B tokens of pre-1931 English. That is a deliberate constraint, not a limitation. The team behind it wants to know if an AI can rediscover the conclusions of past scientists, as Demis Hassabis suggested, or if it is just predicting the next token. We find that question worth taking seriously, and their open-sourced benchmarks and datasets give us all a way to explore it.