feature(sunjx): add rejection sampling in grm_training

[Jiaxuan-Sun]

Rejection Sampling for GRM Training

This directory contains scripts and tools for preparing rejection sampling training data and training GRM (Generative Reward Model) models on both text-to-image (T2I) and text-to-video (T2V) tasks.

Overview

Rejection sampling is a technique to filter high-quality training samples by:

1. Running inference on a dataset using a trained GRM model
2. Filtering correctly predicted samples (where model prediction matches ground truth)
3. Converting filtered samples into training format with Chain-of-Thought (CoT) reasoning
4. Training the model on these high-quality filtered samples