Method pre

data.py:1584–1626

(self, idx)

        return result
    
    def pre(self, idx):
        instruction = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. 

### Instruction:
Analyze the user's interaction history, provide insights about their preferences, and predict the next item's semantic ID.

"""
        tokens = self.tokenizer.encode(instruction, bos=True, eos=False)
        
        row_data = self.data[idx]
        history_and_pref = self.get_history(row_data)
        
        # Skip empty histories or missing targets
        if not history_and_pref['history_semantic_ids'] or not history_and_pref['target_sid']:
            return None
        
        target_output = history_and_pref['output']
        history_and_pref['output'] = ''
        
        prompt = self.generate_prompt(history_and_pref)
        tokens = tokens + self.tokenizer.encode(prompt, bos=False, eos=False)
        attention_mask = [1] * len(tokens)
        
        if self.test:
            return {
                "input_ids": tokens,
                "attention_mask": attention_mask,
            }
        
        golden_tokens = self.tokenizer.encode(target_output, bos=False, eos=True)
        input_prompt_len = len(tokens)
        tokens = tokens + golden_tokens
        attention_mask = [1] * len(tokens)
        labels = [-100] * input_prompt_len + tokens[input_prompt_len:]
        
        if len(tokens) >= self.max_len:
            print(f"Sequence length {len(tokens)} exceeds max_len {self.max_len}")
        
        return {
            "input_ids": tokens[-self.max_len:],
            "attention_mask": attention_mask[-self.max_len:],
            "labels": labels[-self.max_len:],
        }


class UserPreference2sidSFTDataset(BaseDataset):