Chart-to-Text: CNN-RNN vs Vision-Language Models
Published:
Swapped ResNet-50 for EfficientNet-B2 and added dual-LSTM decoding with coverage attention on 27k+ Statista chart images, lifting BLEU-4 from 0.18 to ~0.50.
Posts by Collection
portfolio
Pedagogical Instruction-Tuning for Llama-3.2 — NSF CATpc
Published:
Built a 7.2k high-quality instruction-tuning dataset (Cohen’s κ 0.88) and fine-tuned Llama-3.2 with QLoRA + DPO for the NSF-funded CATpc project — +14% pedagogical alignment, −8% hallucinations vs. baseline.
CultureEval: Quantifying Cultural Alignment in LLMs
Published:
PCA-based evaluation framework over ~97k survey respondents and 96 sociocultural indicators, surfacing systematic underestimation of religious-traditional values in Llama-2, Gemma-3, and Phi-4 (Cohen’s d 0.89–1.17).
LLM-Driven Community Asset Knowledge Graph (Mistral-7B + Bedrock)
Published:
Two-pass LLM extraction pipeline on Amazon Bedrock Agents — Haiku classifier routes documents between Sonnet and a LoRA-fine-tuned Mistral-7B — lifting structured extraction from 0.32 → 0.78 Triple-F1 over 8.3k records, with a React/Cytoscape.js/FastAPI HITL review UI validated at 0.82 Cohen’s κ.
publications
Benchmarking LLMs for Pairwise Causal Discovery in Biomedical and Multi-Domain Contexts
IEEE International Conference on Big Data
Benchmark of 13 open-source LLMs on pairwise causal discovery across 12 datasets — covers both detection and span-level extraction under zero-shot, CoT, and few-shot ICL. Best detection model scores 49.57% (Cdetect); performance collapses on implicit and multi-sentence cases.
Sydney Anuyah, Sneha Shajee Mohan, Bofu Dong, Ankit Singh Chauhan, Sunandan Chakraborty. (2025). "Examining Pairwise Causal Discovery in open-source Large Language Models using Prompt Tuning. " 2025 IEEE International Conference on Big Data. (To appear)
Download Paper
Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities
ACM COMPASS 2026 (Accepted)
Co-created a culturally grounded corpus of insensitive speech with members of Bangladesh’s Hindu and Chakma communities; integrated those narratives into LLM moderation via RAG. Mixed-method evaluation shows RAG-enhanced moderation is more contextually accurate and perceived differently across ethnic lines.
Dipto Das, Achhiya Sultana, Ankit Singh Chauhan, Saadia Binte Alam, Mohammad Shidujaman, Sunandan Chakraborty, and Syed Ishtiaque Ahmed. (2026). "Can LLM-based Content Moderation Identify Insensitive Speech toward Indigenous Ethnic and Religious Minorities?" ACM COMPASS 2026. (Accepted)
Download Paper