Responsible and Fair AI

Cross-lingual Editing in Multilingual Language Models
This research paper introduces the cross-lingual model editing (XME) paradigm, wherein a fact is edited in one language, and the subsequent update propagation is observed across other languages.

Undertsanding Multi-annotation process.
The NLP community has long advocated for the construction of multi-annotator datasets to better capture the nuances of language interpretation, subjectivity, and ambiguity. In this study, we explore on how the performamce of the model varies when trained on dataset with varying annotation budget.

Model Hubs and Beyond
We explore model hubs such as hugging face and ivestigate on how popularity of model is linked to its performance and documentation.

PythonSaga: Redefining the Benchmark to Evaluate Code Generating LLMs
PythonSaga introduces a groundbreaking benchmark for evaluating code-generating LLMs with 185 hand-crafted prompts, balancing 38 programming concepts across difficulty levels to address biases in existing benchmarks.

Poisoning Attacks in Text Classification and Generation
Experimented with clean-label and label-flipping attacks in text generation and classification. Achieving 99% ASR with 95% clean-accuracy on SST-2 for classification. Classification models with triggers: ‘Google’, ‘James Bond’, and ‘cf’. Pretrained GPT-2 with triggers ‘Apple iPhone’: wikitext-2-raw-v1 and wikitext-103-v1.