This research deals with idiomatic expressions, voice (active/passive), speech (direct/indirect), and tense and sentence structures improving the pre-trained MarianMT model for Hindi-to-English translation where issues such as tense errors, pronoun mistakes, and idiom misinterpretation are commonly seen in online translators. Two techniques—layer-wise fine-tuning, where some model layers are tuned in a one-shot fashion at a constant learning rate, and discriminative fine-tuning, which uses different learning rates for different components of the model so that higher layers can be rapidly specialized to the target task while lower layers remain cautiously optimized, were used on this diverse dataset of sentence types and grammar constructs. The layer-wise fine-tuning raises the METEOR to 21.4% and chrF to 19.6%, whereas after discriminative fine-tuning, the METEOR is improved by 11.46% and chrF increases by 17.12% over layer-wise fine-tuning. The results showcase how impactful fine-tuning can be on complex translation problems, offering valuable insights from an industry standpoint toward localization and real-time communication.