Machine learning (ML) has been widely of interest for approximately every task in the legal domain. The foundational ML models lack domain-specific knowledge as they are majorly trained on all the judgments without considering the deeper contextual text specific to a particular category of cases. We hypothesize that the machine learning-trained models can further be of enormous value to the legal community if they are trained on the legal lexical, and that too with the closed lexical (CloLex) categories within the domain. This work aims to analyze the performance of introducing closed lexical categories during the word embedding phase of any prediction task. We present Jud-IPL, an Indian Legal Judgment Prediction Dataset, as a large corpus of 42342 Indian Supreme Court cases annotated for judgment prediction. We also provide a battery of experiments by training several baseline ML models differently, i.e. with mixed judgments and segregating the judgments into civil and criminal categories. The differentiating factor of our work is to emphasize the importance of contextualized word vectors to capture word meanings in context over generic embeddings. Our experiments showed that the CloLex approach to embedding helped the Indian Legal Judgment Prediction models achieve the best scores compared to the state-of-the-art models. We publicly release the Jud-IPL dataset on Kaggle for research and benchmarking purposes. The Jud-IPL word will refer to the following link: https://doi.org/10.34740/kaggle/dsv/14441669 .