Class UserPreference2sidSFTDataset

data.py:1629–1807

class UserPreference2sidSFTDataset(BaseDataset):
    def __init__(self, user_preference_file, index_file, tokenizer, max_len=2048, sample=-1, test=False, seed=0, category="", dedup=False):
        """
        SFT dataset that uses user interaction history with preferences to predict next item's semantic ID.
        Uses interaction history from preference file, predicts the last item in the sequence.
        
        Args:
            user_preference_file: Path to JSON file with user preferences
            index_file: Path to .index.json file mapping item_id to semantic IDs
            tokenizer: Tokenizer for encoding text
            max_len: Maximum sequence length
            sample: Number of samples to use (-1 for all)
            test: Whether this is test mode
            seed: Random seed
            category: Category name for prompts
            dedup: Whether to filter duplicate items
        """
        super().__init__(tokenizer, max_len, test, category, dedup, seed)

        # Load user preferences - handle both JSON and JSONL formats
        with open(user_preference_file, 'r') as f:
            try:
                preference_data = json.load(f)
            except json.JSONDecodeError:
                # Try JSONL format (multiple JSON objects, one per line)
                f.seek(0)
                preference_data = []
                for line in f:
                    line = line.strip()
                    if line:
                        preference_data.append(json.loads(line))
        
        # Handle new flat structure: each item is a separate training sample
        self.training_samples = []
        
        for item in preference_data:
            if item.get('split') == 'train':  # Only process train data
                user_id = item['user']
                preference_text = item.get('user_preference', '')
                context = item.get('context', {})
                history_items = context.get('history_items', [])
                target_item = context.get('target_item')
                
                # Create interaction history by combining history_items and target_item
                interaction_history = history_items + ([target_item] if target_item is not None else [])
                
                # Each item becomes a separate training sample
                self.training_samples.append({
                    'user_id': user_id,
                    'preference_text': preference_text,
                    'interaction_history': interaction_history
                })
        
        # Load index mapping
        with open(index_file, 'r') as f:
            self.indices = json.load(f)
        
        # Prepare training data from preference file interaction histories