trained with verl for paper-query citation chunk grounding.
base model: HuggingFaceTB/SmolLM2-135M-Instruct
dataset: paperbd/paper_preference_150K-v1
training hyperparams: dpo-lr5e-6-ep1-beta0.1-lora16a32-seq1024
local source folder: paperhound
the dataset contains positive cited chunks, not the full arxiv paper haystack, so this model is trained to emit known supporting chunks for a paper/query pair.
