Class SidSFTDataset_GPR

data.py:469–594

class SidSFTDataset_GPR(CSVBaseDataset):
    def __init__(self, train_file, tokenizer, max_len=2048, sample=-1, test=False, seed=0, category="", K=4, dedup=False):
        super().__init__(train_file, sample, seed, max_len, category, dedup, tokenizer, test)

        # Try to load features from standard location
        try:
            with open(f'data/{category}/{category}.user.json', 'r') as f:
                self.user_features = json.load(f)
        except FileNotFoundError:
            try:
                dataset_dir = os.path.dirname(train_file)
                # Assuming structure data/Amazon/train/Sports... -> data/Sports/Sports.user.json
                with open(f'data/{category}/{category}.user.json', 'r') as f:
                    self.user_features = json.load(f)
            except:
                self.user_features = {}
            
        try:
            with open(f'data/{category}/{category}.item.json', 'r') as f:
                self.item_features = json.load(f)
        except FileNotFoundError:
            self.item_features = {}
            
        self.get_inputs()  

    def get_history(self, row):
        row['history_item_sid'] = eval(row['history_item_sid'])
        L = len(row['history_item_sid']) 
        history = ""
        history_str = ", ".join(row["history_item_sid"])
        for i in range(L):
            if i == 0:
                history += row['history_item_sid'][i]
            else:
                history += ", " + row['history_item_sid'][i]      
        target_item = str(row['item_sid'])
        target_item_sid = row["item_sid"]
        last_history_item_sid = row['history_item_sid'][-1] if row['history_item_sid'] else None
        return {"input": f"The user has interacted with items {history} in chronological order. Can you predict the next possible item that the user may expect?",
                "output": target_item + "\n",
                "history_str": history_str,
                "dedup": target_item_sid == last_history_item_sid}
    
    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:
Can you predict the next possible item that the user may expect?

"""
        tokens = self.tokenizer.encode(instruction, bos=True, eos=False)
        
        row = self.data.iloc[idx]
        
        # Heterogeneous Prompt Construction
        user_id = str(row.get('user_id_original_str', ''))
        u_token = self.user_features.get(user_id, '[USER_UNKNOWN]')
        e_token = row.get('e_token', '[CTX_HOMEPAGE]')