Language models built at home match larger models on Brazilian laws
Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese
Computation and LanguageMachine Learning
Summary
Training a language model no longer needs huge data centers. The authors show how to create a Portuguese language model on a consumer GPU using publicly available tools. Their small models perform similarly to larger, more expensive ones on tasks involving Brazilian laws. They tested two methods, autoregressive and masked diffusion, and found that both have strengths depending on the task. This work provides code and benchmarks to help others build specialized language models without huge resources.
What this means in practice
- •For legal technology teams: Build efficient Portuguese language models on modest hardware to detect tampered legal texts and fill document references accurately.
- •For language technology developers: Create specialized language models without data centers using the released Python package and tutorials for Portuguese text tasks.
Authors
Pedro Carvalho Brom
Abstract
Building a language model of one's own no longer requires a data centre. This tutorial trains one from scratch, tokenizer included, on one consumer GPU, and shows how to tell whether it is good for the tasks of a specialist domain. It proposes no new method. It releases a tested Python package and two notebooks in Brazilian Portuguese with autoregressive and masked diffusion generation on one network, a preregistered comparison protocol, legal tasks built from text alone with a Kneser-Ney floor and public yardsticks, reference numbers for models of 11 to 12M parameters, and a catalogue of defects that only measurement revealed. On Brazilian prose and federal laws, at minutes per run, the autoregressive mode peaks at 64 epochs over the same text and then memorises it, while masked diffusion improves up to 256. At their best points on test, the diffusion bound exceeds the autoregressive likelihood by 0.06 bits per byte, within the preregistered minimum effect. On held-out laws and on acts published after training, the model ties a 6-gram floor in likelihood on the same windows, and it loses to the floor at continuing those acts. It beats the floor at detecting tampered text, and fills cross-references more often than the floor and public models of 160M and 630M parameters, only in decrees, a genre absent from training, where those scorers stay near or below chance, and in filling the blank only when each group weighs the same, since per item it ties the 630M model. On laws, the public models, trained with about 5x10^2 to 10^5 times its compute, lead in tampered text and continuation. The blind human reading is preregistered and not part of this version.