Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum

Large language models (LLMs) are commonly trained on datasets consisting of fixed-length token sequences. These datasets are created by randomly concatenating documents of various lengths and then chunking them into sequences of a predetermined target length. However, this method of concatenation can lead to cross-document attention within a sequence, which is neither a desirable learning signal nor computationally efficient. Additionally, training on long sequences becomes computationally prohibitive due to the quadratic cost of attention. In this study, we introduce dataset decomposition, a novel variable sequence length training technique, to tackle these challenges. We decompose a dataset into a union of buckets, each containing sequences of the same size extracted from a unique document. During training, we use variable sequence length and batch size, sampling simultaneously from all buckets with a curriculum. In contrast to the concat-and-chunk baseline, which incurs a fixed attention cost at every step of training, our proposed method incurs a penalty proportional to the actual document lengths at each step, resulting in significant savings in training time. We train an 8k context-length 1B model at the same cost as a 2k context-length model trained with the baseline approach. Experiments on a web-scale corpus demonstrate that our approach significantly enhances performance on standard language evaluations and long-context benchmarks, reaching target accuracy 3x faster compared to the baseline. Our method not only enables efficient pretraining on long sequences but also scales effectively with dataset size. Lastly, we shed light on a critical yet less studied aspect of training large language models: the distribution and curriculum of sequence lengths, which results in a non-negligible difference in performance.

Body Lift and Drag for a Legged Millirobot in Compliant Beam Environment

Much current study of legged locomotion has rightly focused on foot traction forces, including on granular media. Future legged millirobots will need to go through terrain, such as brush or other vegetation, where the body contact forces significantly affect locomotion. In this work, a (previously developed) low-cost 6-axis force/torque sensing shell is used to measure the interaction forces between a hexapedal millirobot and a set of compliant beams, which act as a surrogate for a densely cluttered environment. Experiments with a VelociRoACH robotic platform are used to measure lift and drag forces on the tactile shell, where negative lift forces can increase traction, even while drag forces increase. The drag energy and specific resistance required to pass through dense terrains can be measured. Furthermore, some contact between the robot and the compliant beams can lower specific resistance of locomotion. For small, light-weight legged robots in the beam environment, the body motion depends on both leg-ground and body-beam forces. A shell-shape which reduces drag but increases negative lift, such as the half-ellipsoid used, is suggested to be advantageous for robot locomotion in this type of environment.