Yu-won Lee

Sorry, I made a quick fix for that. It should be at the data_args, but I've made a mistake when refactoring the code.

I'm currently bit busy so, I'm not sure for it. But I think GSPO is supported by trl in default so, you could just add one or 2 line for...

You can add `--importance_sampling_level "sequence_token"` to use GSPO. Also the deafult loss for grpo is DAPO so I think you could just leave it.

I'm not really familiar with the audio type of thing. Need some study for it.

You could see this and make your own dataset. https://github.com/QwenLM/Qwen2.5-VL/blob/main/cookbooks/document_parsing.ipynb It's a bit difference with others when you are using grounding jobs.

Yes, that would be more easy. You could remove this part for sure to pass the original image size. https://github.com/2U1/Qwen2-VL-Finetune/blob/04305d230217a4facd2c54aade95c8a9e4765f4c/src/dataset/data_utils.py#L79-L80

Yes. Grouding task is a bit complicated comparing to other tasks. Maybe I'll make a pipeline for grounding in near future.

You can add a try-except in the dataset something like this. ``` def __getitem__(self, i): try: # the original code except ValueError as e: # log & skip; DataLoader will...

You could make a reward function like this ``` from sacrebleu import corpus_bleu from rouge_score import rouge_scorer _bleu = lambda cand, ref: corpus_bleu(cand, [ref]).score / 100.0 _rouge = rouge_scorer.RougeScorer(["rougeL"], use_stemmer=True)...

1. The format is right. The content betwwen `` and `` is literally the thinking process and answering. You could take a look at other reasoning dataset to make your...