deberta-v3

Efficient and Scalable Estimation of Tool Representations in Vector Space

Finetuning the DeBERTa v3 model for the emotion recognition task

Detect or classify input sentences as grammatically correct or incorrect by fine-tuning pre-trained DeBERTa-v3 model

Application for training the pretrained transformer model DeBERTaV3 on an Aspect Based Sentiment Analysis task

Official repository for the EMNLP 2024 paper "How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics"

Submodular Subset Selection for Long-Document Question Answering

Real-time prompt injection firewall for LLM applications. Detects jailbreaks, prompt leaks, indirect injections, and adversarial prompts.

Finding the source code hidden in the text.

Challenge to distinguish whether a sentence from a news article expresses the subjective view of the author behind it or presents an objective view on the covered topic

Explainable Aspect-Based Sentiment Analysis for Startup Employee Reviews using DeBERTa-v3 and SHAP

Educational workshop for NLP engineers. Fine-tuning DeBERTa-v3 for CEFR level prediction on serverless Modal GPUs.

This project utilises Natural Language Processing to generate questions, create exams and evaluate answers without using an LLM. It uses semantic chunking, retrieval-augmented question/answer generation, and a grading method built on Natural Language Inference and sentence embeddings, grounded in the original source text.

Backbone: 5 x DeBerta and Head: Rapids SVR

Three-class sentiment analysis (satisfied, neutral, not satisfied) on 53k Amazon Echo Dot reviews with a heavy 81/8/11 class imbalance. Compares 7 TF-IDF classical models and 5 fine-tuned transformers (BERT, RoBERTa, DeBERTa and more). Best model: DeBERTa-v3 with 0.752 macro-F1 and 88.7% accuracy. Includes a Flask web app for live predictions.

Real-time prompt injection defense at the edge. A zero-trust Chrome extension and local AI backend that intercepts cognitive threats before they reach your LLMs.

Tactical next-action + reasoning prediction on 348 football match contexts (Shipd Project Eris). 4-component ensemble with task-coupling: DeBERTa-v3-base / large, cross-encoder MCQ scorer, zero-shot NLI, and a three-pass Qwen3.5-35B-A3B-Int4 + Gemma-4-26B-A4B-it MoE fusion with PRM rerank. W&B-instrumented. Target combined ≥ 0.80

Split-based Prompt Injection Detector

RedLockX is an advanced multi-task NLP security model designed to detect:- Prompt Injection Attacks - Jailbreak Attempts - Instruction Overrides - System Prompt Extraction - Role Manipulation - Context Hijacking

IELTS Automated Essay Scoring (AES) – Multi-Task DeBERTa-based Architecture

This repo details code for building a text classifier for predicting Bank Transaction categories. I finetune a base version of a DeBERTaV3 model purely on text data, as well as another version using a combination of text and non-text (e.g., categorical, datetime, etc.) data.