CriticalMaking /Evaluation framework behind Population Fidelity: Python and R pipelines that test whether LLM survey responses reproduce group-level accuracy and the magnitude and structure of between-group variation.
PythonRJupyterLLM Evaluation
neemiasbsilva /Companion training pipeline for the Population Fidelity paper: cultural and distributional LoRA fine-tuning experiments on World Values Survey responses, producing adapters for the evaluation framework to compare.
PythonLoRAFine-tuningPyTorchPEFTMLflowWorld Values Survey
neemiasbsilva /[EMNLP 2026 - PANDORA] Persona Prompting in Multimodal Urban Perception: about 60,000 persona-conditioned captions, perception tags and justifications per model from Qwen3-VL and Gemma-4.
PythonQwen3-VLGemmaPersona Prompting
neemiasbsilva /[DCOSS-IoT 2026] Stable Behavior, Limited Variation: a factorial set of personas instantiated as multimodal LLM agents that judge urban scenes, testing within-persona consistency and cross-persona variation.
PythonLLM AgentsPersona SimulationUrban Computing
neemiasbsilva /Multimodal LLMs see sentiment better through description than by direct classification: 95.8% F1 on PerceptSent, beating CNN and Transformer baselines by up to 15 points, with 72 checkpoints published on Hugging Face.
PythonHugging FaceMultimodal LLMsVisual Sentiment
neemiasbsilva /The notebooks behind the MLLMs book chapter: fundamentals and emblematic models, then preprocessing, prompt engineering and multimodal pipelines built with LangChain and LangGraph.
JupyterLangChainLangGraphMLLMs