教育领域“知之”大模型--山东大学软件学院2024年项目实训（十）

了解history的结构

使大模型融入history,通过浏览下面swift库中的文件可以发现，在编码时，会将query加入到history的末尾，并将query置为空，如果history不传值就会默认为None;所以为了能够利用history,我们只需要传入history的值即可。

    def encode(self, example: Dict[str, Any]) -> Tuple[Dict[str, Any], Dict[str, Any]]:
        """return: inputs, tokenizer_kwargs"""
        if not self._is_init:
            raise ValueError(
                'Template is not initialized, please use the `get_template` function to obtain the template.')
        query: Optional[str] = example.get('query', None)
        response: Optional[str] = example.get('response', None)
        history: Optional[History] = example.get('history', None)
        system: Optional[str] = example.get('system', None)
        if history is None:
            history = []
        if len(history) > 0:
            assert self.support_multi_round, 'The template does not support multi-round chat.'
        if system is None:
            if self.use_default_system:
                system = self.default_system
        elif system == '':
            system = None
        else:
            assert self.prefix_has_system is not None, 'The template does not support `system`.'
        if query is None:
            query = ''
        inputs, tokenizer_kwargs = self._encode(
            query, response, history, system, self.truncation_strategy, auto_add_bos=self.auto_add_bos)
        if inputs.get('labels') is None:
            inputs.pop('loss_scale', None)
        return inputs, tokenizer_kwargs

    def _encode(self,
                query: str,
                response: Optional[str],
                history: History,
                system: Optional[str],
                truncation_strategy: str,
                auto_add_bos: bool = False) -> Tuple[Dict[str, Any], Dict[str, Any]]:
        """
        return: inputs, tokenizer_kwargs
        """
        history = history.copy()
        res_context_list: List[Context] = []
        loss_scale_list: List[float] = []
        if auto_add_bos:
            bos_token_id = self.tokenizer.bos_token_id
            if isinstance(bos_token_id, int) and bos_token_id in self.tokenizer.encode(''):
                res_context_list.append([bos_token_id])
                loss_scale_list.append(0.)
        if system is None:
            prefix = self.prefix
        else:
            prefix = self.prefix_has_system
        self._concat_context_list(prefix, res_context_list, loss_scale_list, system=system)
        history.append([query, response])
        for i, (q, r) in enumerate(history):
            context_list = self.prompt.copy()
            if i < len(history) - 1:
                context_list.append('{{RESPONSE}}')
                context_list += self.chat_sep
            elif r is not None:
                # last response
                context_list.append('{{RESPONSE}}')
                context_list += self.suffix
            if q or r:
                self._concat_context_list(
                    context_list, res_context_list, loss_scale_list, query=q, response=r, round0=i)