train_beauty_sid_rec.py代码中的疑问

[mymymynameisisislry]

作者你好，非常好的工作！
我看到代码中在处理数据时，为了进行label预测，会mask掉一部分数据：

`class CustomDataCollator:
def init(self, tokenizer, mlm=False):
self.tokenizer = tokenizer
self.mlm = mlm

def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
    input_ids = [feature["input_ids"] for feature in features]
    attention_mask = [feature["attention_mask"] for feature in features]

    max_length = max(len(ids) for ids in input_ids)

    padded_input_ids = []
    padded_attention_mask = []
    labels = []

    for i, (ids, mask) in enumerate(zip(input_ids, attention_mask)):
        padding_length = max_length - len(ids)
        padded_ids = ids + [self.tokenizer.pad_token_id] * padding_length
        padded_mask = mask + [0] * padding_length

        label = padded_ids.copy()

        text = self.tokenizer.decode(ids, skip_special_tokens=False)
        user_start_pos = text.find("<|im_start|>user")

        if user_start_pos != -1:
            user_start_tokens = self.tokenizer.encode("<|im_start|>user", add_special_tokens=False)

            for j in range(len(ids) - len(user_start_tokens) + 1):
                if ids[j:j+len(user_start_tokens)] == user_start_tokens:
                    for k in range(j):
                        label[k] = -100
                    break
            else:
                for k in range(len(label)):
                    label[k] = -100
        else:
            for k in range(len(label)):
                label[k] = -100

        padded_input_ids.append(padded_ids)
        padded_attention_mask.append(padded_mask)
        labels.append(label)
    
    return {
        "input_ids": torch.tensor(padded_input_ids, dtype=torch.long),
        "attention_mask": torch.tensor(padded_attention_mask, dtype=torch.long),
        "labels": torch.tensor(labels, dtype=torch.long),
    }`

比如user_start_pos = text.find("<|im_start|>user")
可以理解为mask掉了"<|im_start|>user"之前的system prompt部分吗。如果这样的话，为什么不同样mask掉user prompt部分呢/为什么要让user和assistant部分同时用作label呢？

[zhangrz22]

@mymymynameisisislry
Thank you for your question! The user prompt portion is the user's historical interaction sequence. This sequence contains important collaborative information.
We have the model compute the loss on both the historical interaction sequence and the target because this applies the Causal LM's objective of predicting the next part of a sequence from its previous parts to the entire user behavior sequence, not just to the final target.

[wanxingliu94]

@mymymynameisisislry Thank you for your question! The user prompt portion is the user's historical interaction sequence. This sequence contains important collaborative information. We have the model compute the loss on both the historical interaction sequence and the target because this applies the Causal LM's objective of predicting the next part of a sequence from its previous parts to the entire user behavior sequence, not just to the final target.

This approach seems to be different from one rec v2 where u only train on the last interaction. What’s the reasoning for the difference?

[AgentEvan]

这个自定义的 DataCollator 没有把 label 序列中的 pad_token_id 转换成 -100，实际运行看了下处理后的结果确实是这样，是不是意味着 padding 的 token 也会参与训练？（qwen3 的代码里似乎没找到哪里会显式地将 pad_token_id 转成-100，loss_function 的入参 ignore_index 也只能传一个值）