Deepon Halder

About Me

Research @ AI4Bharat, IIT Madras
UG @ IIEST Shibpur

deeponh.2004@gmail.com
Github
Google Scholar
LinkedIn


I am from India 🇮🇳. I work as a Research Intern at AI4Bharat, IIT Madras under the guidance of Dr. Raj Dabre. I am also known as neuralnets. Feel free to contact me for any queries.

News

2026 2025

Publications

2026

Top-b: Entropic Regulation of Relative Probability Bands in Autoregressive Language Processes

Deepon Halder, Raj Dabre.

Preprint

arXiv

Multilingual TinyStories: A Synthetic Combinatorial Corpus of Indic Children's Stories for Training Small Language Models

Deepon Halder, Angira Mukherjee.

Preprint

arXiv

Not All Time Is Gregorian: Evaluating LLMs on Cultural Calendar Systems

Deepon Halder, Adish Pandya, Raj Dabre.

ICBINB Workshop @ICLR 2026

Scripts Through Time: A Survey of the Evolving Role of Transliteration in NLP

Thanmay Jayakumar, Deepon Halder, Raj Dabre.

ACL Findings 2026

arXiv ProjectPage Github

2025

RiddleBench: A New Generative Reasoning Benchmark for LLMs

Deepon Halder, Alan Saji, Thanmay Jayakumar, Raj Dabre, Ratish Puduppully, Anoop Kunchukuttan.

EACL Findings 2026

arXiv ProjectPage 🤗 HuggingFace

CycleDistill: Bootstrapping Machine Translation using LLMs with Cyclical Distillation

Deepon Halder, Thanmay Jayakumar, Raj Dabre.

MELT Workshop @ COLM 2025 / WAT @ IJCNLP-AACL 2025

arXiv ProjectPage Github

A Comprehensive Survey of Data Poisoning Attacks

Deepon Halder, Anshika Gupta, Diya Ghosh, Hafizur Rehman.

Survey Paper

arXiv

Projects

paper to code : implementation of research papers into code

neugrad : Built a lightweight autograd engine in Python and NumPy, with tensor operations, automatic differentiation, backpropagation, and a scalable design for deep learning, mimicking PyTorch for education and experimentation.

Past Experience

Research Intern : IIT Bhubaneshwar [Dec 2024 - Jan 2025]

Research Apprentice : IIEST Shibpur [July 2024 - Sept 2024]

Events and Achievements over Time

Winner - Cody Challenge 2025 : Finished a month long course on MATLAB, and won the Cody Challenge at the end by solving 20/20 problems in the contest

Winner - Braindead Hackathon : Won 1st prize at this hackathon by building a finetuned Mistral 7B to summarize texts

1st Runners Up - MCode Metallum : Coded up a modern to-do app and won 2nd prize

1st Runners Up - Startup Challenge @ IIEST: Made a project Auto@Kol, which aimed to connect the auto-rickshaw framework efficiently in Kolkata

2nd Runners Up - Hacknovare Instruo : Coded up a mental health management app and won the 3rd prize

Cohere ML School 2025 : Got selected - couldn't attend due to previous engagements

The Residency Feb 2025 : Got selected in the Bangalore Cohort of The Residency - didn't go

KVPY 2021 SA : Extra Merit List Scholar

JBNSTS Scholar : Selected in top 250 students in West Bengal

Technical Blogs

linear regression: A deep dive into Linear Regression and the math behind it.

transformers: a deep dive into transformers and a visual guide into how it works.

optimizers: a deep dive into optimizers and a journey through time.

rnn: a deep dive into recurrent neural networks and how the math behind it works.

basics of nlp [part 1]: discussion into how text preprocessing, regex, frequencies, and word embeddings work.

basics of nlp [part 2]: discussion into how pos tagging, ner, sentiment analysis, and n-gram models work.

basics of nlp [part 3]: a guide into how hidden markov models, text clustering, attention work.

llms [part 1]: a guide into how embeddings, positional embeddings, tokenizer (especially bpe tokenizer) work.

enhance your model [part 1]: a guide into how lora, model distillation, gradient clipping and early stopping work.

llms [part 2]: a guide into how attention works, in great detail.

deepseek r1 explanation: a guide into how deepseek r1 works under the hood.

all about quantization: a guide into how quantization occurs in llms.

Talks & Podcasts

Talks or Podcasts I have done


If you can’t explain something to a first-year student, then you haven’t really understood it.
— R. Feynman