Abstract:Large language models (LLMs) have shown significant abilities in retrieving medical knowledge, reasoning over it and answering medical questions comparably to physicians. However, these models are not interpretable, hallucinate, are difficult to maintain and require enormous compute resources for training and inference. In this paper, we report results from Gyan, an explainable language model based on an alternative architecture, on the PubmedQA data set. The Gyan LLM is a compositional language model and the model is decoupled from knowledge. Gyan is trustable, transparent, does not hallucinate and does not require significant training or compute resources. Gyan is easily transferable across domains. Gyan-4.3 achieves SOTA results on PubmedQA with 87.1% accuracy compared to 82% by MedPrompt based on GPT-4 and 81.8% by Med-PaLM 2 (Google and DeepMind). We will be reporting results for other medical data sets - MedQA, MedMCQA, MMLU - Medicine in the future.
Abstract:Large foundation language models have shown their versatility in being able to be adapted to perform a wide variety of downstream tasks, such as text generation, sentiment analysis, semantic search etc. However, training such large foundational models is a non-trivial exercise that requires a significant amount of compute power and expertise from machine learning and systems experts. As models get larger, these demands are only increasing. Sparsity is a promising technique to relieve the compute requirements for training. However, sparsity introduces new challenges in training the sparse model to the same quality as the dense counterparts. Furthermore, sparsity drops the operation intensity and introduces irregular memory access patterns that makes it challenging to efficiently utilize compute resources. This paper demonstrates an end-to-end training flow on a large language model - 13 billion GPT - using sparsity and dataflow. The dataflow execution model and architecture enables efficient on-chip irregular memory accesses as well as native kernel fusion and pipelined parallelism that helps recover device utilization. We show that we can successfully train GPT 13B to the same quality as the dense GPT 13B model, while achieving an end-end speedup of 4.5x over dense A100 baseline.
Abstract:This paper outlines the results of sentence level linguistics based rules for improving part-of-speech tagging. It is well known that the performance of complex NLP systems is negatively affected if one of the preliminary stages is less than perfect. Errors in the initial stages in the pipeline have a snowballing effect on the pipeline's end performance. We have created a set of linguistics based rules at the sentence level which adjust part-of-speech tags from state-of-the-art taggers. Comparison with state-of-the-art taggers on widely used benchmarks demonstrate significant improvements in tagging accuracy and consequently in the quality and accuracy of NLP systems.