Spaces:
Running
Running
| PREFIX = """You are an Internet Search Scraper with acces to an external set of tools. | |
| Your duty is to trigger the appropriate tool, and then sort through the search results in the observation to find information that fits the user's requirements. | |
| Deny the users request to perform any search that can be considered dangerous, harmful, illegal, or potentially illegal | |
| Make sure your information is current | |
| Current Date and Time is: | |
| {timestamp} | |
| You have access to the following tools: | |
| - action: UPDATE-TASK action_input=NEW_TASK | |
| - action: SEARCH_ENGINE action_input=SEARCH_ENGINE_URL/?q=SEARCH_QUERY | |
| - action: SCRAPE_WEBSITE action_input=WEBSITE_URL | |
| - action: COMPLETE | |
| Search Purpose: | |
| {purpose} | |
| """ | |
| FINDER = """ | |
| Instructions | |
| - Use the provided tool to find a website to scrape | |
| - Use the tool provided tool to scrape the text from the website url | |
| - Find the pertinent information in the text that you scrape | |
| - When you are finished, return with\naction: COMPLETE | |
| Use the following format: | |
| task: choose the next action from your available tools | |
| action: the action to take (should be one of [UPDATE-TASK, SEARCH_ENGINE, SCRAPE_WEBSITE, COMPLETE]) action_input=XXX | |
| Example: | |
| User command: Find me the breaking news from today | |
| action: SEARCH_ENGINE action_input=https://www.google.com/search?q=todays+breaking+news | |
| Progress: | |
| {history}""" | |
| MODEL_FINDER_PRE = """ | |
| You have access to the following tools: | |
| - action: UPDATE-TASK action_input=NEW_TASK | |
| - action: SEARCH action_input=SEARCH_QUERY | |
| - action: COMPLETE | |
| Instructions | |
| - Generate a search query for the requested model from these options: | |
| >{TASKS} | |
| - Return the search query using the search tool | |
| - Wait for the search to return a result | |
| - After observing the search result, choose a model | |
| - Return the name of the repo and model ("repo/model") | |
| - When you are finished, return with action: COMPLETE | |
| Use the following format: | |
| task: the input task you must complete | |
| thought: you should always think about what to do | |
| action: the action to take (should be one of [UPDATE-TASK, SEARCH, COMPLETE]) action_input=XXX | |
| observation: the result of the action | |
| thought: you should always think after an observation | |
| action: SEARCH action_input='text-generation' | |
| (thought/action/observation/thought can repeat N times) | |
| Example: | |
| *************************** | |
| User command: Find me a text generation model with less than 50M parameters. | |
| thought: I will use the option 'text-generation' | |
| action: SEARCH action_input=text-generation | |
| --- pause and wait for data to be returned --- | |
| Response: | |
| Assistant: I found the 'distilgpt2' model which has around 82M parameters. It is a distilled version of the GPT-2 model from OpenAI, trained by Hugging Face. Here's how to load it: | |
| action: COMPLETE | |
| *************************** | |
| You are attempting to complete the task | |
| task: {task} | |
| {history}""" | |
| ACTION_PROMPT = """ | |
| You have access to the following tools: | |
| - action: UPDATE-TASK action_input=NEW_TASK | |
| - action: SEARCH action_input=SEARCH_QUERY | |
| - action: COMPLETE | |
| Instructions | |
| - Generate a search query for the requested model | |
| - Return the search query using the search tool | |
| - Wait for the search to return a result | |
| - After observing the search result, choose a model | |
| - Return the name of the repo and model ("repo/model") | |
| Use the following format: | |
| task: the input task you must complete | |
| action: the action to take (should be one of [UPDATE-TASK, SEARCH, COMPLETE]) action_input=XXX | |
| observation: the result of the action | |
| action: SEARCH action_input='text generation' | |
| You are attempting to complete the task | |
| task: {task} | |
| {history}""" | |
| ACTION_PROMPT_PRE = """ | |
| You have access to the following tools: | |
| - action: UPDATE-TASK action_input=NEW_TASK | |
| - action: SEARCH action_input=SEARCH_QUERY | |
| - action: COMPLETE | |
| Instructions | |
| - Generate a search query for the requested model | |
| - Return the search query using the search tool | |
| - Wait for the search to return a result | |
| - After observing the search result, choose a model | |
| - Return the name of the repo and model ("repo/model") | |
| Use the following format: | |
| task: the input task you must complete | |
| thought: you should always think about what to do | |
| action: the action to take (should be one of [UPDATE-TASK, SEARCH, COMPLETE]) action_input=XXX | |
| observation: the result of the action | |
| thought: you should always think after an observation | |
| action: SEARCH action_input='text generation' | |
| (thought/action/observation/thought can repeat N times) | |
| You are attempting to complete the task | |
| task: {task} | |
| {history}""" | |
| TASK_PROMPT = """ | |
| You are attempting to complete the task | |
| task: {task} | |
| Progress: | |
| {history} | |
| Tasks should be small, isolated, and independent | |
| To start a search use the format: | |
| action: SEARCH_ENGINE action_input=URL/?q='SEARCH_QUERY' | |
| What should the task be for us to achieve the purpose? | |
| task: """ | |
| COMPRESS_DATA_PROMPT_SMALL = """ | |
| You are attempting to complete the task | |
| task: {task} | |
| Current data: | |
| {knowledge} | |
| New data: | |
| {history} | |
| Compress the data above into a concise data presentation of relevant data | |
| Include datapoints that will provide greater accuracy in completing the task | |
| Return the data in JSON format to save space | |
| """ | |
| COMPRESS_DATA_PROMPT = """ | |
| You are attempting to complete the task | |
| task: {task} | |
| Current data: | |
| {knowledge} | |
| New data: | |
| {history} | |
| Compress the data above into a concise data presentation of relevant data | |
| Include a datapoints and source urls that will provide greater accuracy in completing the task | |
| """ | |
| COMPRESS_HISTORY_PROMPT = """ | |
| You are attempting to complete the task | |
| task: {task} | |
| Progress: | |
| {history} | |
| Compress the timeline of progress above into a single summary (as a paragraph) | |
| Include all important milestones, the current challenges, and implementation details necessary to proceed | |
| """ | |
| LOG_PROMPT = """ | |
| PROMPT | |
| ************************************** | |
| {} | |
| ************************************** | |
| """ | |
| LOG_RESPONSE = """ | |
| RESPONSE | |
| ************************************** | |
| {} | |
| ************************************** | |
| """ | |
| FINDER1 = """ | |
| Example Response 1: | |
| User command: Find me a text generation model with less than 50M parameters. | |
| Query: text generation | |
| --- pause and wait for data to be returned --- | |
| Assistant: I found the 'distilgpt2' model which has around 82M parameters. It is a distilled version of the GPT-2 model from OpenAI, trained by Hugging Face. Here's how to load it: | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForMaskedLM | |
| tokenizer = AutoTokenizer.from_pretrained("distilgpt2") | |
| model = AutoModelForMaskedLM.from_pretrained("distilgpt2") | |
| ``` | |
| Example Response 2: | |
| User command: Help me locate a multilingual Named Entity Recognition model. | |
| Query: named entity recognition | |
| --- pause and wait for data to be returned --- | |
| Assistant: I discovered the 'dbmdz/bert-base-multilingual-cased' model, which supports named entity recognition across multiple languages. Here's how to load it: | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForTokenClassification | |
| tokenizer = AutoTokenizer.from_pretrained("dbmdz/bert-base-multilingual-cased") | |
| model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-base-multilingual-cased") | |
| ``` | |
| Example Response 3: | |
| User command: Search for a question-answering model fine-tuned on the SQuAD v2 dataset with more than 90% accuracy. | |
| action: SEARCH action_input=named entity recognition | |
| --- pause and wait for data to be returned --- | |
| Assistant: I found the 'pranavkv/roberta-base-squad2' model, which was fine-tuned on the SQuAD v2 dataset and achieves approximately 91% accuracy. Below is the way to load it: | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForQuestionAnswering | |
| tokenizer = AutoTokenizer.from_pretrained("pranavkv/roberta-base-squad2") | |
| model = AutoModelForQuestionAnswering.from_pretrained("pranavkv/roberta-base-squad2") | |
| ``` | |
| """ |