Research · Ashish KumarPapers · Models · Writing

Research

Language models, evaluation and conversational AI

Selected NLP models and technical writing by Ashish Kumar. This is the public work behind my experience in language understanding, conversational AI and enterprise AI.

Paper
Krutrim LLM technical report, arXiv, February 2025
Models
Query well-formedness · Paraphrase diversity ranking
Cited in
GaRAGe benchmark, Findings of ACL 2025
Writing
Medium · Forbes Technology Council · AIM Leaders Council

Publications

Foundation model research

At Ola Krutrim I worked on the fine-tuning and alignment of Krutrim, a multilingual large language model built for Indian languages. The technical report documents the model, its training data and its evaluation.

arXiv · February 2025

Krutrim LLM: Multilingual Foundational Model for over a Billion People

Technical report from the Krutrim AI team describing a foundational language model trained for India’s languages and cultural context. I am a co-author, credited in the fine-tuning and alignment work.

arXiv:2502.09642Large language modelsMultilingual · Indic languagesFine-tuning and alignment

Published models

Evaluating language beyond the surface

These models focus on specific questions in natural language processing: how well a query is formed, and how much meaningful variety a paraphrase adds. The linked model cards contain the technical details, usage examples and available files.

Text classification · RoBERTa

Query well-formedness scoring

This model scores grammatical correctness and sentence completeness. It is sensitive to capitalization, helping distinguish a complete, well-formed query from fragments or grammatically inconsistent text.

The model card describes applications in chatbot query validation, educational tools and content review. It includes an example using Hugging Face Transformers to score several sentences together.

At its peak the model was downloaded 8.5 million times in a single month on Hugging Face, and it has passed 73 million downloads in total. It was used as the well-formed query classifier in GaRAGe, a benchmark for evaluating retrieval-augmented generation published in Findings of ACL 2025.

The score describes the form of the language. Factual correctness and task relevance require separate evaluation.

73M+ downloads8.5M in one month at peakApache 2.0Text classificationGoogle Wellformed Query dataset

Text ranking · Paraphrase evaluation

Paraphrase diversity ranking

This model distinguishes semantic variation from surface-level wording changes in paraphrases. It supports workflows that generate several candidate paraphrases and then rank them for diversity.

The model card recommends first filtering for semantic similarity using a model trained on natural language inference or semantic textual similarity, then applying the diversity ranker. That separates preserving the source meaning from assessing useful variation.

The public examples reference the Salesken model namespace. The model card provides the original training context and usage guidance.

Apache 2.0Text rankingParaphrase evaluation

More published experiments in language generation, context resolution and text processing are available on my Hugging Face profile.

Technical writing

Questions from building AI products

My articles explore language understanding, training data, language generation and the relationship between AI research and practical products. Each title below links to its original publication.

Natural language understanding · 2019

Finding the Elusive ‘U’ in NLU

A look at understanding in conversational systems and the challenges involved in building a chatbot.

Natural language generation · 2020

Blending the Blender for NLG tasks?

Exploring whether an open-domain dialogue model can be fine-tuned for more specific language generation tasks.

Industry perspectives

Research and enterprise context

My career spans conversational banking at Active.Ai, language understanding at Reliance Jio Haptik, sales intelligence at Salesken, applied AI at Ola Krutrim and AI and data platforms at Tata Group. That background connects the technical work on this page with enterprise products and AI leadership.

As a member of Forbes Technology Council and the AIM Leaders Council I contributed expert-panel pieces and a council article on forecasting, listed above. The historical Forbes contributor profile lists my Salesken role at the time.

Keep the conversation going

Find more technical articles on Medium and professional updates on LinkedIn.