Sai Sundaresan, Harshita Chopra, Atanu R. Sinha, Koustava Goswami, Nagasai Saketh Naidu, Raghav Karan, N Anushka. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Lack of large-scale pretraining data for low resource languages from the Indian sub-continent, leads to their underrepresentation in existing massively multilingual models. In this work, we address this gap by proposing a framework to create large raw audio datasets for such under-represented languages by collating publicly accessible audio content. Leveraging this framework, we present MahaDhwani, a corpus comprising 279K hours of raw audio across 22 Indian languages. To test the utility of MahaDhwani, we pretrain a conformer style model, and then further finetune it to build a multilingual ASR model supporting the 22 languages. Using a hybrid multi-softmax decoder, we balance the benefit of shared parameters which enable crosslingual transfer, and the benefit of dedicated capacity for each language. Our evaluations on the IndicVoices benchmark show the benefits of pre-training, particularly in low-resource settings. We will open-source our framework, code and scripts to reproduce the dataset.

Retrieval-Augmented Generation (RAG) is often used with Large Language Models (LLMs) to infuse domain knowledge or user-specific information. In RAG, given a user query, a retriever extracts chunks of relevant text from a knowledge base. These chunks are sent to an LLM as part of the input prompt. Typically, any given chunk is repeatedly retrieved across user questions. However, currently, for every question, attention layers in LLMs fully compute the Keys and Values (KVs) repeatedly for the input chunks, as state-of-the-art methods cannot reuse KV-caches when chunks appear at arbitrary locations or with arbitrary contexts. Naive reuse leads to output quality degradation. This leads to potentially redundant computations on expensive GPUs and increases latency. In this work, we propose Cache-Craft , a system for managing and reusing precomputed KVs corresponding to the text chunks (which we call chunk-caches ) in RAG-based systems. We present how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache and maintain output quality, and how to efficiently store and evict chunk-caches in the hardware for maximizing reuse while masking any overheads. With real production workloads as well as synthetic datasets, we show that Cache-Craft reduces redundant computation by 51% over SOTA prefix-caching and 75% over full recomputation. Additionally, with continuous batching on a real production workload, we get a 1.6× speed up in throughput for both the LLama-3-8B and 70B models and a 2.1× and 2× reduction in end-to-end response latency respectively, compared to prefix-caching, while maintaining generation quality.