ML Times
May 10, 2025
Vision Now Available in Llama.cpp
- llama.cpp enables multimodal input through the
libmtmdlibrary, supporting tools like llama-mtmd-cli and llama-server for enhanced functionality in AI applications.
- llama.cpp enables multimodal input through the
21 GB/s CSV Parsing Using SIMD on AMD 9950X
- Sep 0.10.0 achieves an impressive 21 GB/s CSV parsing speed on the AMD 9950X, marking a ~3x improvement since its initial release in 2023, driven by optimizations for AVX-512 and enhanced CPU capabilities.
LTXVideo 13B AI video generation
- LTXV 13B is a 13 billion-parameter AI model by Lightricks, achieving 30x faster video generation than its 2B predecessor through advanced multiscale rendering technology.
Adventures in Imbalanced Learning and Class Weight
- Class weighting in imbalanced learning may not significantly enhance model performance, as empirical results suggest that optimal weights are only slightly above unweighted training, contradicting the common practice of using inverse proportion weighting.
[R] The Evolution of RL for Fine-Tuning LLMs (from REINFORCE to VAPO)
- The evolution of reinforcement learning (RL) methods for fine-tuning large language models (LLMs) has progressed from classic techniques like PPO and REINFORCE to advanced methods such as GRPO, ReMax, and VAPO, incorporating innovations like reward shaping and token-level losses.
[P] UQLM: Uncertainty Quantification for Language Models
- UQLM is a Python library that enables zero-resource hallucination detection at generation time, utilizing advanced uncertainty quantification techniques to enhance language model reliability.