Rahul Chalamala  — Research

LoLCATs (Low-rank Linear Conversion via Attention Transfer) is a method to create efficient subquadratic LLMs from existing Transformers. LoLCATs replaces softmax with linear attentions and fine-tunes the model with low-rank adaptation, enabling linear-time and constant-memory generation in open-source LLMs.

RedPajama-V2 is an open dataset for training large language models. The dataset includes over 100B text documents coming from 84 CommonCrawl snapshots and processed using the CCNet pipeline. Out of these, there are 30B documents in the corpus that additionally come with quality signals, and 20B documents that are deduplicated.

LeanDojo is an open-source Lean playground that provides tools, data, models, and benchmarks for LLM-based theorem proving. It features ReProver, a retrieval-augmented prover, and a comprehensive benchmark of 98,734 theorems to facilitate research and development.

The interference of NTS-3 signals with GNSS signals was assessed by developing a Python-based framework to evaluate Spectral Separation Coefficients under ITU-R guidelines. Preliminary findings show minimal interference, and several strategies are proposed to mitigate any potential issues.