Kindly ask for the lack of candidate item for recommendation tasks

[EmeryBAI]

Thanks for your excellent work. I seem to meet a problem when I train the model with
bash ./unimp_task.sh 2e-4 3 4b-instruct all rec 2 2
I find all metrics return 0. After I trace back to the input text, I find that the pipeline.mm_utils.recdataset.process_eval_rec_pair() will only generate the chat history without giving any candidate items, which makes the model impossible to choose the target item.
Here is some related code, (you can see input_seq contains no candidate items)
Would you mind telling me how I can fix the problem.

input_seq = input_seq+f"What is the next item recommended to the user? <answer>"
...
src_text = self.tokenizer(
                input_seq,
                return_tensors="pt",
                add_special_tokens=False,
                truncation=True
            )
        src_item = src_text["input_ids"].squeeze(0)
        src_item_mask = src_text["attention_mask"].squeeze(0)
        
        example = {
            "net_input":{
                "input_ids": src_item,
                "attention_masks": src_item_mask,
                "patch_images": patch_image,
                "input_len": input_len
            },
            "net_output":{
                "output_ids": semantic_id,
            }
        }

        return example