In Part 1 we carved the ConvFinQA corpus into curriculum-aware difficulty tiers through some EDA. Part 2 turned that insight into zero-shot baselines, pitting several LLMs against each tier. The hard bucket still hurt, which is proof that smarter prompting, not bigger models, is our next lever. In this installment, instead of fine-tunes we’ll use DSPy’s programmatic prompt optimizers to lift…
In our previous exploration , we analyzed 3,458 ConvFinQA records and established a curriculum learning framework with three difficulty stages: Easy (≤2 ops, simple context), Medium (2-3 ops, moderate complexity), and Hard (≥4 ops, complex multi-turn reasoning). Now it’s time to put this curriculum to work by implementing and evaluating our models. From Prompting to Programming Traditionally, LLM…
In my current project, I’m exploring financial reasoning over conversational data. With all the buzz around DSPy, I decided it’s the perfect time to dive deep into this tool. Financial reasoning over conversational data presents a unique challenge in NLP: models must not only understand natural language but also perform multi-step numerical computations while maintaining context across dialogue…
In May last year, I spotted a tweet from an Anthropic engineer announcing the BSides CTF: CTF Announcement I’m no CTF expert/practitioner, but I tackled a few during undergrad, with a focus on forensic puzzles. This event mixed steganography with some neural-network trivia, so it sounded like the perfect weekend project. The challenge is available via the Wayback Machine at Anthropic AI Bsides .…
Picking up right where the the last post left off, this follow-up dives into the bread-and-butter building blocks of deep-learning kernels. We’ll implement and benchmark core algorithms-sliding-window pools, tile-wise convolutions, warp-level scans, and more. Puzzle 9: Pooling Pooling is a classic trick in neural networks for shrinking down your data-think of it as a way to “summarize” regions of…
Back on the blog after a long hiatus - this time, I’m shifting gears from just reviewing papers(which are available on my GitHub ) to diving deep into hands-on implementations. I’ve always been interested in systems programming, but somehow never really picked it up. The rate of progress in the GenAI space has been exponential recently, with players like Google [1] reportedly processing 9.7…
Introduction This paper aims to create a framework to map query and doc into semantic vectors via self-attention models. We cant use prior knowledge about important tokens for models based on self-attention. Words are split into different tokens using a tokenization mechanism such as WordPiece. We cannot translate word-level knowledge into different tokens. However, from classical information…
Introduction The NLP world had its ImageNet moment with the introduction of the Transformer in the paper Attention is All you Need . The ability to be able to process multiple words/tokens in parallel and train models without labeled data(using self-attention) led to the creation of multiple models which gave us SOTA results on many interesting tasks such as Question Answering, Summarization, etc.…
Introduction This is the TagLM paper mentioned in Lecture 13 in the CS224N course title Semi-supervised sequence tagging with bidirectional language models This paper demonstrates how we can use context embeddings from BiLSTM models and use it for sequence labelling tasks Paper Introduction Typically, RNN are used only on labelled data to learn the context embeddings of words. Semi supervised…
Introduction This is a new paper which explores the limits of using their new T5 titled How Much Knowledge Can You Pack Into The Parameters of a Language Model? . model in a context-free QA domain. As with the T5 model itself, it is very interesting to see these one-model-to-rule-them-all architectures as they exhibit some form of generalization. I found this paper from Adam Roberts twitter thread…
Introduction This paper review is following the blog from Jay Alammar’s blog on the Illustrated Transformer . The blog can be found here . Paper Introduction New architecture based solely on attention mechanisms called Transformer . Gets rids of recurrent and convolution networks completely. Generally, RNN used to seq-to-seq tasks such as translation, language modelling, etc. Transformer allows…
Introduction REALM is a paper mentioned in the T5 paper titled: How Much Knowledge Can You Pack Into The Parameters of a Language Model? TLDR: This paper retrieves documents that have the information present while solving Question-Answer type problems. NOTE : This post is more like my running notes while reading the paper than a comprehensive blog. I will update this blog once I learn a little…
Introduction Disclaimer This is inspired from Dr. Andrew Gelman’s case study, which can be found here . Specifically: This is heavily inspired by Colin Caroll’s Blog present here . A lot of the plotting code from his blog post has been reused. Josh Duncan’s blog post on the same topic which can be found here . This is not a novel solution. It is merely a replication of Dr. Gelman’s blog in PyMC3.…
Yo! Super excited to finally get my own blog. Hope to write out those long pending articles ASAP now. Stay tuned!