adding benchmaxxer repellant to the open asr leaderboard
hugging face adds private speech datasets to its asr leaderboard to reduce benchmark gaming and provide a more realistic view of model performance across accents and speaking styles.
topic
machine learning
vllm v0 to v1: correctness before corrections in rl
migrating from vllm v0 to v1 for online reinforcement learning required fixing logprob semantics, runtime defaults, weight updates, and fp32 head precision to match training dynamics.
lkv trims llm memory with learned cache budgets
a new method learns optimal key-value cache compression directly from task objectives, improving long-context llm efficiency without heuristic rules.
ratequant tunes kv cache precision for llms
a new method allocates different bit-widths to attention heads in kv cache quantization, avoiding distortion model mismatch to improve large language model serving efficiency.
emo: pretraining mixture of experts for emergent modularity
emo is a mixture-of-experts model trained so that experts self-organize into task-specific groups, allowing strong performance with only a small subset of experts.