Aashish Mahato
Author directory2026
SubVision@WMT26: QLoRA Fine-Tuning of Hy-MT2-1.8B for Chinese to English Video Subtitle Translation
Prajwal Ghimire | Aashish Mahato
Proceedings of the Eleventh Conference on Machine Translation
Prajwal Ghimire | Aashish Mahato
Proceedings of the Eleventh Conference on Machine Translation
We describe \textbf{SubVision}, our submission to the WMT26 Video Subtitle Translation shared task. The system fine-tunes the 1.8B-parameter multilingual translation model Hy-MT2-1.8B \citep{zheng2026hymt2familyfastefficient} using 4-bit QLoRA \citep{dettmers2023qloraefficientfinetuningquantized} on 60,000 sentence pairs sampled from the TVsub Chinese-to-English subtitle corpus \citep{DBLP:journals/corr/abs-1801-03257}. The training pipeline uses a single instruction style prompt template, LoRA adapters (rank 16, $\alpha=32$) on all attention and MLP projections, and early stopping based on dev set sacreBLEU. We fix the decoding with beam search with beam size 4, repetition penalty 1.15, and no-repeat 3-gram constraints. The fine-tuned model on a 200 pair TVsub test split with multiple references achieves sacreBLEU 37.63, chrF 52.76, and TER 53.93, compared with zero-shot NF4 baselines of sacreBLEU 11.60 for Hy-MT2-1.8B and 15.79 for Hy-MT2-7B on the same test pairs, prompt, and decoding configuration. The same configuration is used for official inference, translating subtitle lines independently while preserving SRT timing. We report our full training and inference configuration, describe a data processing detail in parsing the corpus’s multi-document, multi-reference SGM files, and compare our final configuration against nine additional trained configurations that differ in training pair count, sequence-length budget, LoRA rank, and decoding generation length. The paired bootstrap analysis over this comparison shows that given the 200 segment test set size, the nominal ranking of the submitted configuration is not statistically distinguishable from several of the alternatives. We report this analysis along with the point estimates rather than presenting the ranking as an established result. We note that evaluation uses an in-domain TVsub split. The code is available at: \url{https://github.com/praaajg/wmt26-hymt2}.
GABU@WAT 2026: Zero Shot and QLoRA Adaptation of Gemma 4 E4B for Multimodal English to Indic Translation
Shiv Ram Saud | Aashish Mahato | Prajwal Ghimire | Sundeep Dawadi | Ramesh Saud | Sunil Regmi
Proceedings of the Thirteenth Workshop on Asian Translation (WAT 2026)
Shiv Ram Saud | Aashish Mahato | Prajwal Ghimire | Sundeep Dawadi | Ramesh Saud | Sunil Regmi
Proceedings of the Thirteenth Workshop on Asian Translation (WAT 2026)
We describe GABU, our WAT2026 English-to-Indic multimodal translation system (CITATION), which translates image grounded English captions into Hindi, Bengali, Malayalam, and Odia. The task follows the WAT English-to-Indic multimodal track setup with text only, image captioning, and multimodal tracks (CITATION). Our system builds on the pretrained, instruction-tuned vision-language model google/gemma-4-E4B-it. Each input combines a cropped visual region with the English caption, rendered through a chat template with a script constrained translation instruction. We explore two approaches which includes a zero shot baseline and a fine tuned approach of the model using QLORA (rank 64, alpha 128, 4-bit NF4, double quantization), trained for 3,000 steps due to computational resource constraints, reflecting our aim to adapt the model’s existing instruction-following ability rather than train it from scratch. We evaluate both systems using BLEU and CHRF3 under greedy decoding and report macro averaged scores across the target languages. The fine-tuned model substantially outperforms the zero shot baseline and achieves a BLEU score of 43.97 and 69.99 CHRF3 on the test split, and 40.37 BLEU and 68.25 CHRF3 on the challenge split. These results correspond to improvements of 24.39 BLEU and 18.72 CHRF3 points on the test split, and 16.70 BLEU and 11.90 CHRF3 points on the challenge split.
PSGS@2026: From Figures to Citation Graphs: QLoRA Based Multimodal Linking for AstroCLIMB
Prajwal Ghimire | Shiv Ram Saud | Aashish Mahato | Sundeep Dawadi | Ramesh Saud | Sunil Regmi | Manoj Shakya
Proceedings of the 4th Workshop on Artificial Intelligence for Scientific Publications
Prajwal Ghimire | Shiv Ram Saud | Aashish Mahato | Sundeep Dawadi | Ramesh Saud | Sunil Regmi | Manoj Shakya
Proceedings of the 4th Workshop on Artificial Intelligence for Scientific Publications
We describe our system for the AstroCLIMB shared task (WASP 2026), which reconstructs the citation graph of astronomy papers from scientific figures and captions alone. The task requires identifying the relationship between pairs of figures and captions as same_figure, same_paper, related_papers, or unrelated_papers. Our system finetunes google/gemma-4-E4B-it using QLoRA on 8,200 competition training rows with a shared descriptive prompt and greedy decoding. A zero shot baseline achieves 0.4408 macro F1 on a balanced 1,800 pair validation split held out from the competition training file, and the finetuned system substantially improves this citation-relation classification for the challenging related_papers class in particular. Our submitted system achieves 0.64282 F1 on all 10,000 official test pairs. These results show that multimodal finetuning can improve citation-relation inference from scientific figures and captions when document metadata is unavailable.
LINUS@EEUCA 2026: Fine-grained Toxicity Detection in Gaming Chat using Multilingual Transformers
Prajwal Ghimire | Aashish Mahato | Sunil Regmi
Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)
Prajwal Ghimire | Aashish Mahato | Sunil Regmi
Proceedings of the 9th Workshop on Event Extraction and Understanding: Challenges and Applications (EEUCA 2026)
The detection of toxic behavior in online gaming communities is crucial for maintaining safe digital spaces, yet remains challenging due to subtle context-dependent and intent-driven language. The GameTox dataset consists of around 53K World of Tanks chat utterances annotated across six categories: Non-toxic, Insults and Flaming, Other Offensive Texts, Hate and Harassment, Threats, and Extremism (CITATION). Our best performing approach, across multiple transformer-based architecture experimentations, is based on the multilingual BERT variant mmBERT-base fine-tuned with class-weighted cross-entropy loss. The best mmBERT-base model achieved a Macro F1 of 0.5882 during validation and an official test Macro F1 of 0.5104 on the shared task leaderboard. An internal held-out evaluation on a development split yielded 0.4282, which we analyze to understand distributional sensitivity to gaming slang and class imbalance. The code is available at: https://github.com/sunilRegmi-ai/eeuca-toxicity-detection.
Search