Topic
Adapter and LoRA-style training over frozen backbones to reduce what must be exchanged or stored.
The author of a Crimean Tatar recogniser scored it for the first time at 34.6% word error, then found four of his audiobooks sitting in the data twice under filenames a name check had cleared.
Reality
Uncertainty quantification on PDE inverse problems normally costs either an adjoint solve or a large offline training set. A Nature Communications paper reports a sampler whose gradients come from a small local fine-tune instead.
Reality
Jev launched closed on Wednesday, and two days later six teams had published reproductions with almost nothing in common underneath. Only one of them published a score, measured on an eval it curated itself.
Reality
ShadowPEFT is in Hugging Face PEFT's main branch as of a September 15th announcement, and it carries its own hidden state and can be detached as a smaller standalone model. Trying it means installing PEFT from source.
Reality
Three LoRA stages on Qwen2.5-0.5B pushed reward to 1.0 and accuracy below the untrained base. The fix that recovered 43 points was an execution harness that runs both queries and compares the rows.
Reality
Online-SDFT fine-tunes a 230M model on an Android phone from delayed, unlabeled interactions. The teacher is the same frozen base weights with the adapter off, and the write-up reports no measurements.
Reality
NVIDIA's federated learning SDK treats large vision-language updates as a transport problem: externalize big objects, stream tensors, and aggregate against disk instead of server memory.
Reality