arXiv subjects

Ashish Kulkarni

Publications and source records attributed to Ashish Kulkarni.

13 recordsLinked to original sources

arXiv2026

Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems

Designing Optical Character Recognition (OCR) systems for India requires balancing linguistic diversity, document heterogeneity, and deployment constraints. In this paper, we study two training strategies for building multilingual OCR systems with Vision-Language Models through the Chitrapathak series. We first follow a popular multimodal approach, pairing a generic vision encoder with a strong multilingual language model and training the system end-to-end for OCR. Alternatively, we explore fine-tuning an existing OCR model, despite not being trained for the target languages. Through extensive evaluation on multilingual Indic OCR benchmarks and deployment-oriented metrics, we find that the second strategy consistently achieves better accuracy-latency trade-offs. Chitrapathak-2 achieves 3-6x speedup over its predecessor with being state-of-the-art (SOTA) in Telugu (6.69 char ANLS) and second best in the rest. In addition, we present Parichay, an independent OCR model series designed specifically for 9 Indian government documents to extract structured key fields, achieving 89.8% Exact Match score with a faster inference. Together, these systems achieve SOTA performance and provide practical guidance for building production-scale OCR pipelines in the Indian context.

cs.CV↗

arXiv2025

BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages

In the context of pretraining of Large Language Models (LLMs), synthetic data has emerged as an alternative for generating high-quality pretraining data at scale. This is particularly beneficial in low-resource language settings where the benefits of recent LLMs have been unevenly distributed across languages. In this work, we present a systematic study on the generation and evaluation of synthetic multilingual pretraining data for Indic languages, where we construct a large-scale synthetic dataset BhashaKritika, comprising 540B tokens using 5 different techniques for 10 languages. We explore the impact of grounding generation in documents, personas, and topics. We analyze how language choice, both in the prompt instructions and document grounding, affects data quality, and we compare translations of English content with native generation in Indic languages. To support scalable and language-sensitive evaluation, we introduce a modular quality evaluation pipeline that integrates script and language detection, metadata consistency checks, n-gram repetition analysis, and perplexity-based filtering using KenLM models. Our framework enables robust quality control across diverse scripts and linguistic contexts. Empirical results through model runs reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

cs.CL↗

arXiv2025

MUTANT: A Recipe for Multilingual Tokenizer Design

Tokenizers play a crucial role in determining the performance, training efficiency, and the inference cost of Large Language Models (LLMs). Designing effective tokenizers for multilingual LLMs is particularly challenging due to diverse scripts and rich morphological variation. While subword methods like Byte Pair Encoding (BPE) are widely adopted, their effectiveness in multilingual settings remains underexplored. We present MUTANT, a recipe for building multilingual tokenizers, with careful vocabulary and training data design, language-aware pre-tokenization, and subword and multiword aware training. We also introduce MUTANT-Indic, a tokenizer for India-specific multilingual LLMs, that produces linguistically coherent tokens and achieves state-of-the-art performance. Evaluated across English, 22 Indian languages and code data, our tokenizer improves the average fertility score by 39.5%$ over LLaMA4 and by 18% over Sutra (the current best). This translates to 44% improvement in inference throughput over LLaMA4 while maintaining comparable performance on English and Indic benchmarks. We present detailed ablations across tokenizer training data size, vocabulary size, merging techniques, and pre-tokenization strategies, demonstrating the robustness of our design choices.

cs.CL↗

arXiv2025

VoiceAgentBench: Are Voice Assistants ready for agentic tasks?

Large scale Speech Language Models have enabled voice assistants capable of understanding natural spoken queries and performing complex tasks. However, existing speech benchmarks largely focus on isolated capabilities such as transcription or question answering and do not systematically evaluate agentic behavior or adversarial robustness. To address this, we introduce VoiceAgentBench, a comprehensive benchmark for evaluating SpeechLMs in realistic spoken agentic settings, comprising 6,000+ synthetic spoken queries spanning single-tool invocations, multi-tool workflows, multi-turn dialogue, and safety evaluations across English and six Indic languages. To ensure speaker diversity, we further simulate speaker variability using a novel sampling strategy that selects audios for TTS voice conversion based on speaker embeddings to maximize acoustic diversity. Our evaluation measures tool selection accuracy, structural consistency, and the correctness of tool invocations, including adversarial robustness. Across agentic tasks, ASR-LLM pipelines outperform end-to-end SpeechLMs, achieving up to 60.6% average parameter-filling accuracy on English, while SpeechLMs exhibit lower performance and sharper degradation on Indic languages. All models struggle in sequential workflows and safety evaluations, highlighting persistent limitations in tool orchestration, multilingual generalization, and safety robustness. VoiceAgentBench is publicly available on Hugging Face at https://huggingface.co/datasets/krutrim-ai-labs/VoiceAgentBench, and the codebase is released at https://github.com/ola-krutrim/VoiceAgentBench.

cs.AI↗

arXiv2025

Pragyaan: Designing and Curating High-Quality Cultural Post-Training Datasets for Indian Languages

The effectiveness of Large Language Models (LLMs) depends heavily on the availability of high-quality post-training data, particularly instruction-tuning and preference-based examples. Existing open-source datasets, however, often lack multilingual coverage, cultural grounding, and suffer from task diversity gaps that are especially pronounced for Indian languages. We introduce a human-in-the-loop pipeline that combines translations with synthetic expansion to produce reliable and diverse Indic post-training data. Using this pipeline, we curate two datasets: Pragyaan-IT (22.5K) and Pragyaan-Align (100K) across 10 Indian languages covering 13 broad and 56 sub-categories, leveraging 57 diverse datasets. Our dataset protocol incorporates several often-overlooked dimensions and emphasize task diversity, multi-turn dialogue, instruction fidelity, safety alignment, and preservation of cultural nuance, providing a foundation for more inclusive and effective multilingual LLMs.

cs.CL↗

arXiv2025

Chitranuvad: Adapting Multi-Lingual LLMs for Multimodal Translation

In this work, we provide the system description of our submission as part of the English to Lowres Multimodal Translation Task at the Workshop on Asian Translation (WAT2024). We introduce Chitranuvad, a multimodal model that effectively integrates Multilingual LLM and a vision module for Multimodal Translation. Our method uses a ViT image encoder to extract visual representations as visual token embeddings which are projected to the LLM space by an adapter layer and generates translation in an autoregressive fashion. We participated in all the three tracks (Image Captioning, Text only and Multimodal translation tasks) for Indic languages (ie. English translation to Hindi, Bengali and Malyalam) and achieved SOTA results for Hindi in all of them on the Challenge set while remaining competitive for the other languages in the shared task.

cs.CL↗

arXiv2024

MATHSENSEI: A Tool-Augmented Large Language Model for Mathematical Reasoning

Tool-augmented Large Language Models (TALMs) are known to enhance the skillset of large language models (LLMs), thereby, leading to their improved reasoning abilities across many tasks. While, TALMs have been successfully employed in different question-answering benchmarks, their efficacy on complex mathematical reasoning benchmarks, and the potential complementary benefits offered by tools for knowledge retrieval and mathematical equation solving are open research questions. In this work, we present MathSensei, a tool-augmented large language model for mathematical reasoning. We study the complementary benefits of the tools - knowledge retriever (Bing Web Search), program generator + executor (Python), and symbolic equation solver (Wolfram-Alpha API) through evaluations on mathematical reasoning datasets. We perform exhaustive ablations on MATH, a popular dataset for evaluating mathematical reasoning on diverse mathematical disciplines. We also conduct experiments involving well-known tool planners to study the impact of tool sequencing on the model performance. MathSensei achieves 13.5% better accuracy over gpt-3.5-turbo with Chain-of-Thought on the MATH dataset. We further observe that TALMs are not as effective for simpler math word problems (in GSM-8K), and the benefit increases as the complexity and required knowledge increases (progressively over AQuA, MMLU-Math, and higher level complex questions in MATH). The code and data are available at https://github.com/Debrup-61/MathSensei.

cs.CL↗

arXiv2023

Resolving the Hydrophobicity of Me-4PACz Hole Transport Layer for High-Efficiency Inverted Perovskite Solar Cells

[4-(3,6-dimethyl-9H-carbazole-9-yl)butyl]phosphonic acid (Me-4PACz) self-assembled monolayer (SAM) has been employed in perovskite single junction and tandem devices demonstrating high efficiencies. However, a uniform perovskite layer does not form due to the hydrophobicity of Me-4PACz. Here, we tackle this challenge by adding a conjugated polyelectrolyte poly(9,9-bis(3'-(N,N-dimethyl)-N-ethylammonium-propyl-2,7-fluorene)-alt-2,7-(9,9dioctylfluorene)dibromide (PFN-Br) to the Me-4PACz in a specific ratio, defines as Pz:PFN. With this mixing engineering strategy of Pz:PFN, the PFN-Br interacts with the A-site cation and is confirmed via solution-state nuclear magnetic resonance studies. The narrow full width at half maximum (FWHM) of diffraction peaks of perovskite film revealed improved crystallization on the optimal mixing ratio of Pz:PFN. Interestingly, the mixing of PFN-Br additionally tunes the work function of the Me-4PACz as revealed by the Kelvin probe force microscopy and built-in-voltage estimation in solar cells. Devices employing optimized Pz:PFN mixing ratio deliver open-circuit voltage (Voc)of 1.16 V and efficiency >20% for perovskites with a bandgap of 1.6 eV with high reproducibility and concomitant stability. Considering significant research on Me-4PACz SAM, our work highlights the importance of obtaining a uniform perovskite layer with improved yield and performance.

cond-mat.mtrl-sci↗

arXiv2023

MFBE: Leveraging Multi-Field Information of FAQs for Efficient Dense Retrieval

In the domain of question-answering in NLP, the retrieval of Frequently Asked Questions (FAQ) is an important sub-area which is well researched and has been worked upon for many languages. Here, in response to a user query, a retrieval system typically returns the relevant FAQs from a knowledge-base. The efficacy of such a system depends on its ability to establish semantic match between the query and the FAQs in real-time. The task becomes challenging due to the inherent lexical gap between queries and FAQs, lack of sufficient context in FAQ titles, scarcity of labeled data and high retrieval latency. In this work, we propose a bi-encoder-based query-FAQ matching model that leverages multiple combinations of FAQ fields (like, question, answer, and category) both during model training and inference. Our proposed Multi-Field Bi-Encoder (MFBE) model benefits from the additional context resulting from multiple FAQ fields and performs well even with minimal labeled data. We empirically support this claim through experiments on proprietary as well as open-source public datasets in both unsupervised and supervised settings. Our model achieves around 27% and 20% better top-1 accuracy for the FAQ retrieval task on internal and open datasets, respectively over the best performing baseline.

cs.IR↗

arXiv2023

Characterizing the Influence of Charge Extraction Layers on the Performance of Triple-Cation Perovskite Solar Cells

Selecting suitable charge transport layers and suppressing non-radiative recombination at interfaces to the absorber layer are vital to maximize the efficiency of halide perovskite solar cells. In this work, high-quality perovskite thin films and devices are fabricated with different fullerene-based electron transport layers and different self-assembled monolayers as hole transport layers. We then perform a comparative study of a significant variety of different electrical, optical and photoemission-based characterization techniques to quantify the properties of the solar cells, the individual layers and importantly the interfaces between them. In addition, we highlight the limitations and problems of the different measurements, the insights gained by combining different methods and the different strategies to extract information from the experimental raw data.

cond-mat.mtrl-sci↗

arXiv2023

A Universal Strategy of Perovskite Ink-Substrate Interaction to Overcome the Poor Wettability of a Self-Assembled Monolayer for Reproducible Perovskite Solar Cells

Perovskite solar cells employing self assembled monolayers such as Me-4PACz as hole transport layer has been reported to demonstrate high device efficiency. However, the poor perovskite wetting on the Me-4PACz caused by poor perovskite ink interaction with the underlying Me-4PACz presents significant challenges for fabricating efficient perovskite devices. A triple co-solvent system comprising of dimethylformamide (DMF), dimethyl sulfoxide (DMSO) and N-methyl-2-pyrrolidone (NMP) is employed to improve the perovskite ink-substrate interaction and obtain a uniform perovskite layer. In comparison to DMF, DMSO-based inks, the inclusion of NMP shows considerably higher binding energies of the perovskite ink with Me-4PACz as revealed by density-functional theory calculations. With the optimized triple co-solvent ratio, the perovskite devices deliver high power conversion efficiencies of >20%, 19.5% and ~18.5% for active areas of 0.16 cm2, 0.72 cm2 and 1.08 cm2 respectively. Importantly, this perovskite ink-substrate interaction approach is universal and helps in obtaining a uniform layer and high photovoltaic device performance for other perovskite compositions such as MAPbI3, FAMAPbI3-xBrx, and MA-free FACsPbI3-xBrx.

physics.app-ph↗

arXiv2022

Improving Speech Prosody of Audiobook Text-to-Speech Synthesis with Acoustic and Textual Contexts

We present a multi-speaker Japanese audiobook text-to-speech (TTS) system that leverages multimodal context information of preceding acoustic context and bilateral textual context to improve the prosody of synthetic speech. Previous work either uses unilateral or single-modality context, which does not fully represent the context information. The proposed method uses an acoustic context encoder and a textual context encoder to aggregate context information and feeds it to the TTS model, which enables the model to predict context-dependent prosody. We conducted comprehensive objective and subjective evaluations on a multi-speaker Japanese audiobook dataset. Experimental results demonstrate that the proposed method significantly outperforms two previous works. Additionally, we present insights about the different choices of context - modalities, lateral information and length - for audiobook TTS that have never been discussed in the literature before.

cs.SD↗

arXiv2022

Extraction of Product Specifications from the Web -- Going Beyond Tables and Lists

E-commerce product pages on the web often present product specification data in structured tabular blocks. Extraction of these product attribute-value specifications has benefited applications like product catalogue curation, search, question answering, and others. However, across different Websites, there is a wide variety of HTML elements (like , , , , etc.) typically used to render these blocks that makes their automatic extraction a challenge. Most of the current research has focused on extracting product specifications from tables and lists and, therefore, suffers from recall when applied to a large-scale extraction setting. In this paper, we present a product specification extraction approach that goes beyond tables or lists and generalizes across the diverse HTML elements used for rendering specification blocks. Using a combination of hand-coded features and deep learned spatial and token features, we first identify the specification blocks on a product page. We then extract the product attribute-value pairs from these blocks following an approach inspired by wrapper induction. We created a labeled dataset of product specifications extracted from 14,111 diverse specification blocks taken from a range of different product websites. Our experiments show the efficacy of our approach compared to the current specification extraction models and support our claim about its application to large-scale product specification extraction.

cs.IR↗