Projects
Distributed LLM Inference Service
An efficient cloud-based inference service for LLMs using Kubernetes. Leveraging vLLM, an open-source library for optimizing LLMs, the service addresses challenges in memory consumption and latency.
RLHF on Microsoft/Phi-2 for dialogues
Fine-tuning Phi-2 using Direct Preference Optimization (DPO) on the Anthropic Helpful and Harmless dialogue dataset, leveraging LoRA adapters and 8-bit quantization for efficient training, showing improved reward margins.
Conditional Diffusion Model for Next Frame Generation
A diffusion model conditioned on past 11 frames for auto-regressive generation of next 11 frames. Achieved an MSE of 0.004 for next frame prediction and Jaccard score of 30.89 for semantic segmentation.
transformerkp: A transformers based library for keyphrase identification from text documents.
transformerkp allows you to train and apply state-of-the-art deep learning models for keyphrase extraction and generation from text documents. It supports several benchmark datasets and evaluation metrics for keyphrase extraction and generation.
t-CRF: CRF head on top of Transformer for sequence tagging.
t-CRF enables users to use Conditional Random Field layer on top of any Transformer based sequence tagger like POS tagger, entity recognition, etc.
SpanElectra: A language model with accuracy of spanBERT and efficiency of ELECTRA.
SpanElectra is an efficient language model that uses span boundary objective from SpanBert LM to capture span-level context but uses the discriminator-generator training method inspired by Electra LM for efficient low-resource training.
Question-Generation using Language Model.
Given a paragraph, generate all the possible questions related to this paragraph. If a context is provided, create those questions from that paragraph whose answer should be that context.