Lillama: Large Language Models Compression via Low-Rank Feature Distillation

Current LLM structured pruning methods typically involve two steps: (1) compression with calibration data and (2) costly continued pretraining on billions of tokens to recover lost performance. This second step is necessary as the first significantly impacts model accuracy. Prior research suggests pretrained Transformer weights aren't inherently low-rank, unlike their activations, which may explain this drop. Based on this observation, we propose Lillama, a compression method that locally distills activations with low-rank weights. Using SVD for initialization and a joint loss combining teacher and student activations, we accelerate convergence and reduce memory use with local gradient updates. Lillama compresses Mixtral-8x7B within minutes on a single A100 GPU, removing 10 billion parameters while retaining over 95% of its original performance. Phi-2 3B can be compressed by 40% with just 13 million calibration tokens, resulting in a small model that competes with recent models of similar size. The method generalizes well to non-transformer architectures, compressing Mamba-3B by 20% while maintaining 99% performance.

Large Language Models Compression via Low-Rank Feature Distillation

Current LLM structured pruning methods involve two steps: (1) compressing with calibration data and (2) continued pretraining on billions of tokens to recover the lost performance. This costly second step is needed as the first step significantly impacts performance. Previous studies have found that pretrained Transformer weights aren't inherently low-rank, unlike their activations, which may explain this performance drop. Based on this observation, we introduce a one-shot compression method that locally distills low-rank weights. We accelerate convergence by initializing the low-rank weights with SVD and using a joint loss that combines teacher and student activations. We reduce memory requirements by applying local gradient updates only. Our approach can compress Mixtral-8x7B within minutes on a single A100 GPU, removing 10 billion parameters while maintaining over 95% of the original performance. Phi-2 3B can be compressed by 40% using only 13 million calibration tokens into a small model that competes with recent models of similar size. We show our method generalizes well to non-transformer architectures: Mamba-3B can be compressed by 20% while maintaining 99% of its performance.

BabySLM: language-acquisition-friendly benchmark of self-supervised spoken language models

Self-supervised techniques for learning speech representations have been shown to develop linguistic competence from exposure to speech without the need for human labels. In order to fully realize the potential of these approaches and further our understanding of how infants learn language, simulations must closely emulate real-life situations by training on developmentally plausible corpora and benchmarking against appropriate test sets. To this end, we propose a language-acquisition-friendly benchmark to probe spoken language models at the lexical and syntactic levels, both of which are compatible with the vocabulary typical of children's language experiences. This paper introduces the benchmark and summarizes a range of experiments showing its usefulness. In addition, we highlight two exciting challenges that need to be addressed for further progress: bridging the gap between text and speech and between clean speech and in-the-wild speech.