SphereLab · Research blog
Rotating Sphere.
Research blogs on the principles that we are chasing to make scale tractable.
-
Understanding Parameter-Efficient Finetuning from a Stability-Plasticity PerspectiveA stability-plasticity benchmark and geometry-based diagnosis of parameter-efficient finetuning on LLMs.
-
Orbit: A Ultra-efficient RL Post-training Pipeline for Trillion-Parameter LLMsA deployment-aligned, low-precision, PEFT-centric RL pipeline that stably and efficiently perform 1T-parameter LLM post-training on a single GPU node with 8×B200.
-
Pion: A Stable Optimizer for LLMsA new geometric route to stable LLM training.