Dataset for Finetuning VLM & Running Inferences

This dataset supports the development and evaluation of the FLARE framework for detecting transient flares in quasar light curves. It is intended for parameter-efficient fine-tuning and benchmarking of Vision Language Models (VLMs) on morphology-based classification tasks in irregularly sampled astrophysical time series.

The dataset consists of fully simulated quasar light curves generated from a Damped Random Walk (DRW) process, calibrated to reproduce the cadence, noise characteristics, and variability properties of SDSS Stripe 82 observations. Synthetic transient events are injected in flux space to produce physically consistent magnitude variations.

The dataset comprises five classes: (i) baseline DRW variability (non-flare), (ii) FRED (Fast Rise Exponential Decay) flares, (iii) Gaussian flares, (iv) Gamma-profile flares, and (v) single-epoch spike artifacts. Event parameters, including amplitude and temporal scales, are sampled across physically motivated ranges to ensure diversity in morphology.

Each sample is provided as a rendered light curve image, designed to preserve temporal structure while enabling direct ingestion by VLMs. Accompanying JSONL files define structured prompts and ground-truth labels for supervised fine-tuning and evaluation. The test split is additionally used for cross-model benchmarking and inference.

This dataset is designed for research in time-domain astronomy, statistical anomaly detection, and machine learning methods applied to astrophysical variability. It is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license to support reproducible and extensible research.