Method format

ppo/ppo_datahelper.py:314–335

(self, sample: Tuple[List[str], str])

        return output

    def format(self, sample: Tuple[List[str], str]) -> Dict[str, Any]:
        # original text concat special prompt: human prompt and assistant prompt
        context = [get_special_prompt(i, self.opt) + u for i, u in enumerate(sample)]
            
        context_vec = self.tokenizer.encode(
            self.tokenizer.eos_token.join(context) + self.tokenizer.eos_token,
            add_special_tokens=True
        )
        
        text_vec = context_vec[:self.opt.maxlen_prompt]
        loss_mask = []
        cnt = 0
        for v in text_vec:
            loss_mask.append(cnt % 2)
            cnt += int(v == self.tokenizer.eos_token_id)

        output = {
            'text_vec': text_vec,
            'loss_mask': loss_mask,
        }

        return output

    def batchify(self, batch_samples: List[Dict[str, Any]]) -> Dict[str, Any]:
        batch = dict()