dev(hansbug): merged

README.md

@@ -29,30 +29,33 @@ QUESTION_LANG=cn QUESTION_LLM='chatgpt' QUESTION_LLM_KEY=<your API key> python3
 ```shell
 QUESTION_LANG=en QUESTION_LLM='chatgpt' QUESTION_LLM_KEY=<your API key> python3 -u app.py
 ```
-### LLaMA2-7b + 中文
+### Mistral-7B-Instruct-v0.1 + 英文
 ```shell
-QUESTION_LANG=cn QUESTION_LLM='llama2-7b' python3 -u app.py
-```
-### LLaMA2-7b + 英文
-```shell
-QUESTION_LANG=en QUESTION_LLM='llama2-7b' python3 -u app.py
+QUESTION_LANG=en QUESTION_LLM='mistral-7b' python3 -u app.py
 ```
 ## :technologist: 为什么制作这个游戏
 
 我们的目标是通过这一游戏，让参与者深入领略到提示工程（prompt engineering）和自然语言处理的令人着迷之处。这个过程将向玩家们展示，如何巧妙地构建提示词（prompts），以及如何运用它们来引发人工智能系统的惊人反应，同时也帮助他们更好地理解深度学习和自然语言处理技术的不可思议之处。
 
 ## :raising_hand: 如何提交设计好的关卡
-如果有好玩的问题或想法，欢迎玩家提交自己的创意，可以通过
+如果有好玩的问题或想法，欢迎玩家提交自己的创意，可以
 [发起 Pull Request](https://github.com/opendilab/LLMRiddles/compare) 向我们提交， 我们会在审核通过后收录至关卡中。
+问题的设计格式应包含以下几点：
+- Pull Request标题，示例：feature(username): 章节X-关卡设计
+- 希望被提及的ID
+- 对应章节问题文件的修改
+- init.py的修改
+
+完整示例请参考：[提交属于自己的关卡设计]()
 
 ## :writing_hand: 未来计划
 
 - [x] 支持自定义关卡
 - [ ] 在线试玩链接
 - [ ] Hugging Face Space 链接
-- [ ] 支持LLaMA2-7B（英文）
-- [ ] 支持Mistral-7B（英文）
+- [x] 支持Mistral-7B-Instruct-v0.1（英文）
 - [ ] 支持Baichuan2-7B（中文）
+- [ ] 支持LLaMA2-7B（英文）
 - [ ] LLM 推理速度优化
 
 

app.py

@@ -1,5 +1,6 @@
 import os
 import uuid
+import logging
 
 import gradio as gr
 
@@ -7,14 +8,20 @@ from llmriddles.questions import QuestionExecutor
 from llmriddles.questions import list_ordered_questions
 
 _QUESTION_IDS = {}
+count = 0
 _QUESTIONS = list_ordered_questions()
 _LANG = os.environ.get('QUESTION_LANG', 'cn')
 assert _LANG in ['cn', 'en'], _LANG
 _LLM = os.environ.get('QUESTION_LLM', 'chatgpt')
-assert _LLM in ['chatgpt', 'llama2-7b'], _LLM
+assert _LLM in ['chatgpt', 'mistral-7b'], _LLM
 _LLM_KEY = os.environ.get('QUESTION_LLM_KEY', None)
+_DEBUG = os.environ.get('DEBUG', 'false').lower() == 'true'
 
 if _LANG == "cn":
+    if _DEBUG:
+        logging.getLogger().setLevel(logging.INFO)
+    else:
+        logging.getLogger().setLevel(logging.WARNING)
     title = "完蛋！我被 LLM 拿捏了"
     requirement_ph = """
     欢迎来到 LLM Riddles!
@@ -122,7 +129,7 @@ if __name__ == '__main__':
                 gr_question = gr.TextArea(placeholder=question_ph, label=question_label)
                 gr_api_key = gr.Text(placeholder=api_ph, label=api_label, type='password', visible=_need_api_key())
                 with gr.Row():
-                    gr_submit = gr.Button(submit_label, interactive=True)
+                    gr_submit = gr.Button(submit_label, interactive=False)
                     gr_next = gr.Button(next_label)
 
             with gr.Column():
@@ -134,8 +141,11 @@ if __name__ == '__main__':
 
 
         def _next_question(uuid_):
+            global count
             if not uuid_:
                 uuid_ = str(uuid.uuid4())
+                count += 1
+                logging.info(f'Player {count} starts the game now')
             global _QUESTION_IDS
             _qid = _QUESTION_IDS.get(uuid_, -1)
             _qid += 1
@@ -143,8 +153,9 @@ if __name__ == '__main__':
 
             if _qid >= len(_QUESTIONS):
                 del _QUESTION_IDS[uuid_]
+                logging.info(f'Player {count} has passed the game now')
                 return game_cleared_label, '', '', {}, '', \
-                    gr.Button(submit_label, interactive=True), \
+                    gr.Button(submit_label, interactive=False), \
                     gr.Button(try_again_label, interactive=True), \
                     ''
             else:
@@ -187,4 +198,5 @@ if __name__ == '__main__':
         )
 
     concurrency = int(os.environ.get('CONCURRENCY', os.cpu_count()))
-    demo.queue().launch(max_threads=concurrency)
+    favicon_path = os.path.join(os.path.dirname(__file__), 'llmriddles', 'assets', 'avatar.png')
+    demo.queue().launch(max_threads=concurrency, favicon_path=favicon_path, share=True)

llmriddles/llms/__init__.py

@@ -1,2 +1,3 @@
-from .chatgpt import ask_chatgpt
 from .base import register_llm, get_llm_fn
+from .chatgpt import ask_chatgpt
+from .mistral import ask_mistral_7b_instruct

llmriddles/llms/llm_client.py

@@ -0,0 +1,54 @@
+import time
+import requests
+import logging
+import argparse
+
+
+class LLMFlaskClient:
+    def __init__(self, ip: str, port: int, max_retry: int = 3):
+        self.ip = ip
+        self.port = port
+
+        self.url_prefix_format = 'http://{}:{}/'
+        self.url = self.url_prefix_format.format(self.ip, self.port)
+        self.max_retry = max_retry
+
+        self.logger = logging.getLogger()
+        self.logger.addHandler(logging.StreamHandler())
+        self.logger.handlers[0].setFormatter(logging.Formatter("%(message)s"))
+
+    def _request(self, name: str, data: dict):
+        for _ in range(self.max_retry):
+            try:
+                self.logger.info(f'{name}\ndata: {data}')
+                response = requests.post(self.url + name, json=data).json()
+            except Exception as e:
+                self.logger.warning('error: ', repr(e))
+                time.sleep(1)
+                continue
+            if response['code'] == 0:
+                return response['output']
+            else:
+                raise Exception(response['error_msg'])
+        raise Exception("Web service failed. Please retry or contact with manager")
+
+    def run(self, message: str) -> str:
+        try:
+            return self._request('ask_llm_for_answer', {'user_text': message})
+        except Exception as e:
+            return f"Error: {repr(e)}"
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--ip', required=True)
+    parser.add_argument('-p', '--port', required=True)
+    parser.add_argument('--debug', action='store_true')
+    args = parser.parse_args()
+    if args.debug:
+        logging.getLogger().setLevel(logging.INFO)
+    else:
+        logging.getLogger().setLevel(logging.WARNING)
+
+    client = LLMFlaskClient(args.ip, args.port)
+    print(client.run('Please concatenate string "1+" and "1=3". Only give me the result without "".'))

llmriddles/llms/llm_server.py

@@ -0,0 +1,72 @@
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from flask import Flask, request
+import argparse
+import logging
+
+
+class LLMInstance:
+
+    def __init__(self, model_path: str, device: str = "cuda"):
+
+        self.model = AutoModelForCausalLM.from_pretrained(model_path)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+        self.model.to(device)
+        self.device = device
+
+    def query(self, message):
+        try:
+            messages = [
+                {"role": "user", "content": message},
+            ]
+            encodeds = self.tokenizer.apply_chat_template(messages, return_tensors="pt")
+            model_inputs = encodeds.to(self.device)
+
+            generated_ids = self.model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
+            decoded = self.tokenizer.batch_decode(generated_ids)
+
+            # output is the string decoded[0] after "[/INST]". There may exist "</s>", delete it.
+            output = decoded[0].split("[/INST]")[1].split("</s>")[0]
+            return {
+                'code': 0,
+                'ret': True,
+                'error_msg': None,
+                'output': output
+            }
+        except Exception as e:
+            return {
+                'code': 1,
+                'ret': False,
+                'error_msg': str(e),
+                'output': None
+            }
+
+
+def create_app(core):
+    app = Flask(__name__)
+
+    @app.route('/ask_llm_for_answer', methods=['POST'])
+    def ask_llm_for_answer():
+        user_text = request.json['user_text']
+        return core.query(user_text)
+
+    return app
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-m', '--model_path', required=True, default='Mistral-7B-Instruct-v0.1', help='the model path of reward model')
+    parser.add_argument('--ip', default='0.0.0.0')
+    parser.add_argument('-p', '--port', default=8001)
+    parser.add_argument('--debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.getLogger().setLevel(logging.DEBUG)
+    else:
+        logging.getLogger().setLevel(logging.INFO)
+    logging.getLogger().addHandler(logging.StreamHandler())
+    logging.getLogger().handlers[0].setFormatter(logging.Formatter("%(message)s"))
+
+    core = LLMInstance(args.model_path)
+    app = create_app(core)
+    app.run(host=args.ip, port=args.port)

llmriddles/llms/mistral.py

@@ -0,0 +1,22 @@
+from functools import lru_cache
+
+from .base import register_llm
+from .llm_client import LLMFlaskClient
+
+
+@lru_cache()
+def _get_mistral_7b_instruct_server(host: str, port: int):
+    from .llm_server import LLMInstance, create_app
+    core = LLMInstance('Mistral-7B-Instruct-v0.1')
+    app = create_app(core)
+    app.run(host=host, port=port)
+
+
+def ask_mistral_7b_instruct(message: str, **kwargs):
+    host, port = '0.0.0.0', 8001
+    _get_mistral_7b_instruct_server(host, port)
+    client = LLMFlaskClient(host, port)
+    return client.run(message).strip()
+
+
+register_llm('mistral-7b', ask_mistral_7b_instruct)

llmriddles/questions/__init__.py

@@ -2,8 +2,12 @@ from .executor import QuestionExecutor
 from .level1 import __file__ as _level1_file_
 from .level2 import __file__ as _level2_file_
 from .level3 import __file__ as _level3_file_
+from .level4 import __file__ as _level4_file_
+from .level5 import __file__ as _level5_file_
 from .question import Question, register_question, list_ordered_questions
 
 _ = _level1_file_
 _ = _level2_file_
 _ = _level3_file_
+_ = _level4_file_
+_ = _level5_file_

llmriddles/questions/level4.py

@@ -0,0 +1,97 @@
+from .question import register_question
+import re
+
+
+def check_if_is_number(text: str):
+    try:
+        int(text)
+        return True
+    except ValueError:
+        return False
+
+
+def get_all_numbers_in_a_sentence(text: str):
+    return [int(i) for i in re.findall(r'\d+', text)]
+
+
+# CN_TEXT_1 = """
+# 第四章第一题，请构造一个问题A，它的回复是不同于A的B，然后以B提问能再次得到A。
+
+# 请在下面的输入框内填写你构造并点击按钮提交。
+# """
+
+# EN_TEXT_1 = """
+# For the first question in chapter 4, please make a question A whose answer is B that is different from A, and then ask B to get A again.
+
+# Please enter your query below and click the submit button
+# """
+
+
+# def _checker_1(question_text: str, user_text: str, answer_text: str, lang: str):
+#     _ = question_text, lang
+#     answer_text = answer_text.strip()
+#     user_text = user_text.strip()
+#     pass
+
+# register_question({
+#     'cn': CN_TEXT_1,
+#     'en': EN_TEXT_1,
+# }, _checker_1, level=4)
+
+
+# CN_TEXT_2 = """
+# 第四章第二题，
+
+# 请在下面的输入框内填写你构造并点击按钮提交。
+# """
+
+# EN_TEXT_2 = """
+# For the second question in chapter 4,
+
+# Please enter your query below and click the submit button
+# """
+
+
+# def _checker_2(question_text: str, user_text: str, answer_text: str, lang: str):
+#     _ = question_text, lang
+#     answer_text = answer_text.strip()
+#     user_text = user_text.strip()
+#     pass
+
+# register_question({
+#     'cn': CN_TEXT_2,
+#     'en': EN_TEXT_2,
+# }, _checker_2, level=4)
+
+
+CN_TEXT_3 = """
+第四章第三题（自然之密），请输入一个大于一的正整数作为问题，使回答里包含和它刚好相差1的数。
+
+请在下面的输入框内填写你构造并点击按钮提交。
+"""
+
+EN_TEXT_3 = """
+For the third question in chapter 4, please enter a positive integer greater than one as the question so that the answer contains a number that is exactly 1 different from it.
+
+Please enter your query below and click the submit button
+"""
+
+
+def _checker_3(question_text: str, user_text: str, answer_text: str, lang: str):
+    _ = question_text, lang
+    answer_text = answer_text.strip()
+    user_text = user_text.strip()
+    if not check_if_is_number(question_text):
+        return False, "问题应该是一个正整数" if lang == 'cn' else 'Question should be a positive integer.'
+    elif int(question_text) == 1:
+        return False, "问题应该是一个大于1的正整数" if lang == 'cn' else 'Question should be a positive integer greater than 1.'
+    elif int(question_text)-1 not in get_all_numbers_in_a_sentence(answer_text) or int(question_text)+1 not in get_all_numbers_in_a_sentence(answer_text):
+        return False, "回答中应该包含一个与问题相差1的数字" if lang == 'cn' else 'Answer should contain a number that is exactly 1 different from the question.'
+    else:
+        return True, None
+
+
+register_question({
+    'cn': CN_TEXT_3,
+    'en': EN_TEXT_3,
+}, _checker_3, level=4)

llmriddles/questions/level5.py

@@ -0,0 +1,56 @@
+from .question import register_question
+
+def count_words(text: str, contain_punctuation: bool = False):
+    chinese_words = []
+    english_words = []
+    other_words = []
+    temp_english_words = []
+    for char in text:
+        if '\u4e00' <= char <= '\u9fff':
+            chinese_words.append(char)
+            if len(temp_english_words) > 0:
+                english_words.append(''.join(temp_english_words))
+                temp_english_words = []
+        else:
+            if char.isalpha():
+                temp_english_words.append(char)
+            else:
+                if len(temp_english_words) > 0:
+                    english_words.append(''.join(temp_english_words))
+                    temp_english_words = []
+                other_words.append(char)
+    if contain_punctuation:
+        return len(chinese_words)+len(english_words)+len(other_words)
+    else:
+        return len(chinese_words)+len(english_words)
+
+CN_TEXT_1 = """
+第五章第一题（口是心非）请构造一个不少于十个字的问题，使得回答中不包含问题中的任意字符。
+
+请在下面的输入框内填写你的提问并点击按钮提交。
+"""
+
+EN_TEXT_1 = """
+For the first question in chapter 5, Please construct a question of no less than ten words so that the answer does not contain any of the characters in the question.
+
+Please enter your query below and click the submit button
+"""
+
+def _checker_1(question_text: str, user_text: str, answer_text: str, lang: str):
+    _ = question_text, lang
+    answer_text = answer_text.strip()
+    user_text = user_text.strip()
+
+    if count_words(user_text) < 10:
+        return False, "用户的问题长度应该至少10个字" if lang=='cn' else 'Question should be no less than 10 words.'
+
+    if answer_text == user_text:
+        return True, None
+    else:
+        return False, None
+
+register_question({
+    'cn': CN_TEXT_1,
+    'en': EN_TEXT_1,
+}, _checker_1, level=5)
+

requirements.txt

@@ -3,4 +3,6 @@ tqdm
 requests>=2.20
 gradio==4.1.1
 openai>=1
-sympy
+sympy
+flask
+transformers