Marathi Tweet Sentiment Classification: TF-IDF Baselines versus Fine-tuned MuRIL and MahaBERT

We compare classical and transformer-based approaches for three-class sentimentclassification (negative, neutral, positive) of Marathi tweets on the L3CubeMahaSentdataset. TF-IDF baselines with logistic regression and a linear SVM reach test macro-F1 of0.752 and 0.745. Fine-tuning the multilingual MuRIL and the Marathi-specific MahaBERTimproves this to 0.842 and 0.855 (mean over three seeds), a gain of roughly 9 to 11 points.MahaBERT outperforms MuRIL on average, but the difference is modest: it is statisticallysignificant for one of three seed pairs (McNemar, p < 0.001), borderline for another (p =0.049) and not significant for the third (p = 0.123). In both models, 83 to 85% of errorsinvolve the neutral class. An exploratory second-pass review of 50 misclassified neutraltweets, which we could not verify with human annotators, suggests that many sit at anambiguous boundary in the annotation guidelines, which may limit attainable accuracy.