SimPO: Simple Preference Optimization with a Reference-Free Reward
Paper • 2405.14734 • Published • 13
Fit-only OCR preference experiment
Experimental SFT-anchored, length-normalized preference training. Only authored code and owned model/checkpoint artifacts are public. Competition images, labels, generated negatives, predictions and row membership are excluded. The original competition dataset and private owner split are required to reproduce. No leaderboard improvement is claimed.
Method: https://arxiv.org/abs/2405.14734 Initialization: hassanql/road-qwen3vl8b-stackmix@15d83cc593a4e7a61fe58e5e545c75f31e05ddba Base: Qwen/Qwen3-VL-8B-Instruct@0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for hassanql/road-visual-preference-fit
Base model
Qwen/Qwen3-VL-8B-Instruct