--- title: "Mano_sub@DravidianLangTech 2026: Article-Aware Batching" description: "The paper addresses the DravidianLangTech 2026 shared task on Telugu Prompt-Style Recovery. 2 references." authors: ["Manohar Sita Rama Madhurapantula", "Seshu Babu Pulagara"] year: 2026 venue: "Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages" source: "https://aclanthology.org/2026.dravidianlangtech-1.45/" doi: "10.18653/v1/2026.dravidianlangtech-1.45" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/madhurapantula-2026-mano-sub-dravidianlangtech-2026-article/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/madhurapantula-2026-mano-sub-dravidianlangtech-2026-article/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/35/) / Mano\_sub@DravidianLangTech 2026: Article-Aware Batching and Discriminative Fine-Tuning of MuRIL for Telugu Prompt-Style Classification # Mano\_sub@DravidianLangTech 2026: Article-Aware Batching and Discriminative Fine-Tuning of MuRIL for Telugu Prompt-Style Classification **Authors:** Manohar Sita Rama Madhurapantula and Seshu Babu Pulagara · **Year:** 2026 · **Venue:** [Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages](https://md.research.elamur.ai/free/en/venues/dravidianlangtech/) ([Workshop on Speech, Vision, and Language Technologies for Dravidian Languages 2026](https://md.research.elamur.ai/free/en/venues/dravidianlangtech/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.dravidianlangtech-1.45/) · [PDF](https://aclanthology.org/2026.dravidianlangtech-1.45.pdf) · [DOI 10.18653/v1/2026.dravidianlangtech-1.45](https://doi.org/10.18653/v1/2026.dravidianlangtech-1.45) · [OpenAlex](https://openalex.org/W7166853317) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The paper addresses the DravidianLangTech 2026 shared task on Telugu Prompt-Style Recovery. The goal is to assign Telugu text to one of nine styles, such as Formal, Humorous, Serious or Persuasive. The authors found a structural problem in the dataset: its 384 source articles each appear about 7.8 times in different style rewrites. With ordinary random batching, rewrites of the same article often land in one mini-batch. Because MuRIL encodes them almost identically, training collapses to predicting a single class (macro F1 of 0.022). The fix is an article-aware sampler that fills batches round-robin across articles, combined with full fine-tuning of google/muril-base-cased using layer-specific learning rates. Five-fold cross-validation gave a mean macro F1 of 0.3834. The corrected model scored 0.2765 on the official test set. That is far above the team's submitted 0.0491, which ranked 12th, and would have placed 2nd of 13 teams. The study shows that duplicated-source structure in a dataset can silently break training, and that a simple sampling change can repair it. The authors note that part of the gain also came from longer training, which they did not separate in an ablation. ## Key points - **Data:** 3,464 samples in total, with 3,166 used for cross-validation and a 301-sample official test set. - **Setup:** MuRIL with 512-token inputs, batch size 8, AdamW, 10 epochs, trained on CPU. - **Learning rates:** 1e-3 for the classifier head, 3e-5 for the top three layers plus pooler, and 1e-5 for the remaining layers. - **Development results:** five-fold CV mean macro F1 was 0.3834 (std 0.0189), with fold bests between 0.3488 and 0.4040, after 85.74 hours of CPU training. All nine classes were being predicted by epoch 5. - **Test results:** the fold-1 checkpoint reached macro F1 0.2765, compared with 0.2987 for the winning team Error_500. - **Per-class behaviour:** Serious was the hardest class (test F1 0.03), mostly misread as Formal or Authoritative. Pessimistic and Humorous were the easiest on the development set. - **Resources:** code at github.com/msrmanohar/ACL-PRLLM and data on Hugging Face (msrmanohar/telugu-prompt-style-recovery). *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions Also on the byline: [National Institute of Technical Teachers Training and Research, Chennai](https://ror.org/04rxw8s48). 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works ## Cite Cite as: Manohar Sita Rama Madhurapantula and Seshu Babu Pulagara. 2026. Mano\_sub@DravidianLangTech 2026: Article-Aware Batching and Discriminative Fine-Tuning of MuRIL for Telugu Prompt-Style Classification. Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{madhurapantula2026mano, title = {{Mano\_sub@DravidianLangTech 2026: Article-Aware Batching and Discriminative Fine-Tuning of MuRIL for Telugu Prompt-Style Classification}}, author = {Manohar Sita Rama Madhurapantula and Seshu Babu Pulagara}, booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages}, year = {2026}, doi = {10.18653/v1/2026.dravidianlangtech-1.45}, url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.45} } ``` - [← previous Mango: Multi-Agent Web Navigation via Global-View Optimization](https://md.research.elamur.ai/free/en/papers/tong-2026-mango-multi-agent-web-navigation/) - [next → Map-of-Actions: Deliberate Reasoning over Multi-Labeled Graphs](https://md.research.elamur.ai/free/en/papers/ni-2026-map-actions-deliberate-reasoning-over/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/35/) ## References 2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Premjith B, Jyothish Lal G, Bharathi Raja Chakravarthi, Saranya Rajiakodi, Thenmozhi Durairaj, Ratnavel Rajalakshmi, Rahul Ponnusamy, and Chinthala Bhuvanesh. 2026. Shared Task on Prompt Recovery for LLM in Telugu. In Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for… [find](https://scholar.google.com/scholar?q=Premjith+B%2C+Jyothish+Lal+G%2C+Bharathi+Raja+Chakravarthi%2C+Saranya+Rajiakodi%2C+Thenmozhi+Durairaj%2C+Ratnavel+Rajalakshmi%2C+Rahul+Ponnusamy%2C+and+Chinthala+Bhuvanesh%2E+2026%2E+Shared+Task+on+Prompt+Recovery+for+) 2. A. Bala. 2025. Fake news detection in Dravidian languages using MuRIL. Master’s thesis, IIIT Hyderabad. \#\# Limitations [find](https://scholar.google.com/scholar?q=A%2E+Bala%2E+2025%2E+Fake+news+detection+in+Dravidian+languages+using+MuRIL%2E+Master%E2%80%99s+thesis%2C+IIIT+Hyderabad%2E+%5C%23%5C%23+Limitations) *Updated 1 November 2026*