a

app.py

@@ -91,7 +91,7 @@ with gr.Blocks(title="Automatic Prompt Engineer", theme=gr.themes.Glass()) as de
     gr.Markdown('--------')
     gr.Markdown('\n\n')
     gr.Markdown("# SELF INSTRUCT")
-    gr.Markdown('## 第一步：输入参数并上传数据')
+    gr.Markdown('## 第一步：确认参数并上传种子指令')
     with gr.Row():
         with gr.Column():
             n_human = gr.Slider(label="人工指令数", minimum=1, maximum=5, step=1, value=2)
@@ -107,7 +107,7 @@ with gr.Blocks(title="Automatic Prompt Engineer", theme=gr.themes.Glass()) as de
             self_instance = gr.State()
 
     gr.Markdown('\n\n')
-    gr.Markdown('## 第二步：采样few-shot并生成，每点一次会重采样并生成，生成结果会累计')
+    gr.Markdown('## 第二步：采样并生成新指令，每点一次会重采样并生成，生成结果会累计')
     with gr.Row():
         with gr.Column(scale=1):
             fewshot = gr.Textbox(label='采样few-shot')

self/generate.py

@@ -47,7 +47,7 @@ class SELF(object):
         self.scorer = None  # rougeL用于文本相似度计算
         self.all_instruction_tokens = []  # 全部指令，用于新指令消重
         self.all_instruction = []  # 全部指令，用于新指令消重
-        self.sample_few_shot = None
+        self.sample_few_shot = None #每轮few-shot采样填充成prompt，for gradio
         self.load_seed_task(seed_file)
         self.init(prompt)
 
@@ -87,12 +87,11 @@ class SELF(object):
         seed_sample = random.sample(self.human_instruction_data, self.n_human)
         machine_sample = random.sample(self.machine_instruction_data,
                                        min(self.n_machine, len(self.machine_instruction_data)))
-        self.sample_few_shot = seed_sample + machine_sample
         # build few-shot
-        few_shot = gen_few_shot_prompt(self.sample_few_shot)
-        few_shot += SELF.prefix.format(id=self.first_id)  # 新生成的指令
+        self.sample_few_shot = gen_few_shot_prompt(seed_sample + machine_sample)
         # generate
-        result = self.chain({'few_shot': few_shot, 'n_instruct': self.n_instruct})
+        result = self.chain({'few_shot': self.sample_few_shot + SELF.prefix.format(id=self.first_id), # 待生成指令id
+                             'n_instruct': self.n_instruct})
         return result
 
     def decode_response(self, response):
@@ -184,6 +183,6 @@ def init_instance(seed_file, openai_key, n_human, n_machine, n_instruct, prompt)
 def generate_instruction(self_instance):
     keep_instruct_data = self_instance.step()
 
-    return (json.dumps(self_instance.sample_few_shot, ensure_ascii=False),
+    return (self_instance.sample_few_shot,
             json.dumps(keep_instruct_data, ensure_ascii=False),
             f'已生成{self_instance.n_gen} 可用{self_instance.n_keep}')