import os import re import copy import time import gradio as gr from text_generation import Client from transformers import load_tool from share_btn import community_icon_html, loading_icon_html, share_js, share_btn_css HF_TOKEN = os.environ.get("HF_TOKEN", None) os.environ["HF_ALLOW_CODE_EVAL"] = "1" print(HF_TOKEN) FIM_PREFIX = "" FIM_MIDDLE = "" FIM_SUFFIX = "" FIM_INDICATOR = "" theme = gr.themes.Monochrome( primary_hue="indigo", secondary_hue="blue", neutral_hue="slate", radius_size=gr.themes.sizes.radius_sm, font=[ gr.themes.GoogleFont("Open Sans"), "ui-sans-serif", "system-ui", "sans-serif", ], ) tool = load_tool("vwxyzjn/pyserini-wikipedia-kilt-doc") tool_fn = lambda x: tool(x).split("\n")[1][:600] # limit the amount if token, system_prompts clients = { "StarCoderBase TriviaQA": [ Client( "https://api-inference.huggingface.co/models/vwxyzjn/starcoderbase-triviaqa", headers={"Authorization": f"Bearer {HF_TOKEN}"}, ), {"Wiki": tool_fn}, """\ Answer the following question: Q: In which branch of the arts is Patricia Neary famous? A: Ballets A2: Patricia NearyPatricia Neary (born October 27, 1942) is an American ballerina, choreographer and ballet director, who has been particularly active in Switzerland. She has also been a highly successful ambassador for the Balanchine Trust, bringing George Balanchine's ballets to 60 cities around the globe. Result=Ballets Q: Who won Super Bowl XX? A: Chicago Bears A2: Super Bowl XXSuper Bowl XX was an American football game between the National Football Conference (NFC) champion Chicago Bears and the American Football Conference (AFC) champion New England Patriots to decide the National Football League (NFL) champion for the 1985 season. The Bears defeated the Patriots by the score of 46–10, capturing their first NFL championship (and Chicago's first overall sports victory) since 1963, three years prior to the birth of the Super Bowl. Super Bowl XX was played on January 26, 1986 at the Louisiana Superdome in New Orleans. Result=Chicago Bears """, ["Q: In which country is Oberhofen situated?", "Q: Irish Olympic champion Michelle smith was suspended in 1999 over drug allegations in which sport?"] ], "StarCoderBase GSM8K": [ Client( "https://api-inference.huggingface.co/models/lvwerra/starcoderbase-gsm8k", headers={"Authorization": f"Bearer {HF_TOKEN}"}, ), {"PythonInterpreter": load_tool("lvwerra/python-interpreter")}, """\ Example of using a Python API to solve math questions. Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left? def solution(): money_initial = 23 bagels = 5 bagel_cost = 3 money_spent = bagels * bagel_cost money_left = money_initial - money_spent result = money_left return result print(solution()) 72 Result = 72 """, ["Q: Tim has $400, and he received $1021. How much does he have?"] ], } def parse_tool_call(text, request_token="", call_token=""): """ Parse request string. Expected format: query """ result = re.search(f"(?<={request_token}).*?(?={call_token})", text, re.DOTALL) # if we can't find a / span we return none if result is None: return None, None else: extracted_text = result.group() result = re.search(r"<(.*?)>", extracted_text) # if we can't find a tool name we return none if result is None: return None, None else: tool = result.group(1) # split off the tool name query = ">".join(extracted_text.split(">")[1:]) return tool, query def generate( prompt, system_prompt, version, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0, ): client, tools, _, _ = clients[version] temperature = float(temperature) if temperature < 1e-2: temperature = 1e-2 top_p = float(top_p) fim_mode = False # TextEnv tool generate_kwargs = dict( temperature=temperature, max_new_tokens=max_new_tokens, top_p=top_p, repetition_penalty=repetition_penalty, do_sample=True, seed=42, stop_sequences=["", ""] ) generation_still_running = True request_idx = -1 call_idx = -1 response_idx = -1 submit_idx = -1 i = 0 while generation_still_running: try: stream = client.generate_stream(system_prompt + prompt, **generate_kwargs) # call env phase output = system_prompt + prompt generation_start_idx = len(output) highlighted_output = [ (prompt, "query"), ] yield highlighted_output, output[generation_start_idx:] for response in stream: i += 1 output += response.token.text tool, query = parse_tool_call(output[generation_start_idx:]) if tool is not None and query is not None: # print("=====tool", i, tool, response, output) if tool not in tools: response = f"Unknown tool {tool}." try: response = tools[tool](query) output += response + "" except Exception as error: response = f"Tool error: {str(error)}" if request_idx == -1: request_idx = output[generation_start_idx:].find("") if call_idx == -1: call_idx = output[generation_start_idx:].find("") if call_idx != -1: call_idx += len("") if response_idx == -1: response_idx = output[generation_start_idx:].find("") if response_idx != -1: response_idx += len("") if submit_idx == -1: submit_idx = output[generation_start_idx:].find("") # I am sorry about the code print("-------", generation_start_idx, request_idx, call_idx, response_idx) highlighted_output = [ (prompt, "query"), (output[generation_start_idx:], "model") if request_idx == -1 else ("", ""), (output[generation_start_idx:generation_start_idx+request_idx], "model"), (output[generation_start_idx+request_idx:], "model") if call_idx == -1 else ("", ""), (output[generation_start_idx+request_idx:generation_start_idx+call_idx], "tool request"), (output[generation_start_idx+call_idx:generation_start_idx+response_idx], "tool call"), (output[generation_start_idx+response_idx:], "model") if submit_idx != -1 else ("", ""), # (output[generation_start_idx:generation_start_idx+request_idx], ""), # (output[generation_start_idx+request_idx:generation_start_idx+call_idx], "request"), # (output[generation_start_idx+call_idx:], "call"), ] print(i, highlighted_output, output[generation_start_idx:]) yield highlighted_output, output[generation_start_idx:] # breakpoint() call_output = copy.deepcopy(output) print("start submit output") # response phase generate_kwargs["stop_sequences"] = [""] stream = client.generate_stream(output, **generate_kwargs) for response in stream: output += response.token.text if submit_idx == -1: submit_idx = output[generation_start_idx:].find("") # print("-------", generation_start_idx, request_idx, call_idx, response_idx) highlighted_output = [ (prompt, "query"), (output[generation_start_idx:generation_start_idx+request_idx], "model"), (output[generation_start_idx+request_idx:generation_start_idx+call_idx], "request"), (output[generation_start_idx+call_idx:generation_start_idx+response_idx], "call"), (output[generation_start_idx+response_idx:], "model") if submit_idx != -1 else ("", ""), ] # print(highlighted_output, output[generation_start_idx:]) yield highlighted_output, output[generation_start_idx:] print("-------", generation_start_idx, request_idx, call_idx, response_idx) print(highlighted_output, output[generation_start_idx:]) return highlighted_output, output[generation_start_idx:] except Exception as e: if "loading" in str(e): gr.Warning("waiting for model to load... (this could take up to 20 minutes, after which things are much faster)") time.sleep(7) continue else: raise gr.Error(str(e)) examples = [ "X_train, y_train, X_test, y_test = train_test_split(X, y, test_size=0.1)\n\n# Train a logistic regression model, predict the labels on the test set and compute the accuracy score", "// Returns every other value in the array as a new array.\nfunction everyOther(arr) {", "Poor English: She no went to the market. Corrected English:", "def alternating(list1, list2):\n results = []\n for i in range(min(len(list1), len(list2))):\n results.append(list1[i])\n results.append(list2[i])\n if len(list1) > len(list2):\n \n else:\n results.extend(list2[i+1:])\n return results", ] def process_example(args): for x in generate(args): pass return x css = ".generating {visibility: hidden}" monospace_css = """ #q-input textarea { font-family: monospace, 'Consolas', Courier, monospace; } """ css += share_btn_css + monospace_css + ".gradio-container {color: black}" description = """

This is a demo to generate text the following StarCoderBase models fine-tuned using TRL's TextEnvironment:

  • StarCoderBase TriviaQA: Uses a Wikipedia search index to answer trivia questions. It was trained on the TriviaQA dataset.
  • StarCoderBase GSM8K: Uses a Python Interpreter to answer math questions. It was trained on the GSM8K dataset.
""" with gr.Blocks(theme=theme, analytics_enabled=False, css=css) as demo: with gr.Column(): gr.Markdown(description) with gr.Row(): version = gr.Dropdown( list(clients.keys()), value=list(clients.keys())[0], label="Model", info="Choose a model from the list", ) with gr.Row(): with gr.Column(): instruction = gr.Textbox( value="Q: In which country is Oberhofen situated?", # placeholder="Enter your question here. E.g., Q: In which country is Oberhofen situated?", lines=2, label="Input", ) submit = gr.Button("Generate", variant="primary") output = gr.HighlightedText( label="Output", color_map={"query": "red", "tool call": "green", "tool response": "blue", "model": "pink"}, ) gr_examples = gr.Examples( examples=[example for client in clients.values() for example in client[3]], inputs=[instruction], cache_examples=False, ) with gr.Row(): with gr.Column(): with gr.Accordion("Raw output", open=False): output2 = gr.Code(elem_id="q-output", lines=30, label="Raw output") with gr.Accordion("Advanced settings", open=False): with gr.Row(): column_1, column_2 = gr.Column(), gr.Column() with column_1: temperature = gr.Slider( label="Temperature", value=0.2, minimum=0.0, maximum=1.0, step=0.05, interactive=True, info="Higher values produce more diverse outputs", ) max_new_tokens = gr.Slider( label="Max new tokens", value=256, minimum=0, maximum=8192, step=64, interactive=True, info="The maximum numbers of new tokens", ) with column_2: top_p = gr.Slider( label="Top-p (nucleus sampling)", value=0.90, minimum=0.0, maximum=1, step=0.05, interactive=True, info="Higher values sample more low-probability tokens", ) repetition_penalty = gr.Slider( label="Repetition penalty", value=1.2, minimum=1.0, maximum=2.0, step=0.05, interactive=True, info="Penalize repeated tokens", ) with gr.Accordion("Prompt", open=False): system_prompt = gr.Textbox( value=clients[list(clients.keys())[0]][2], label="System prompt", ) version.select( lambda x: (clients[x][2]), inputs=[version], outputs=[system_prompt], ) submit.click( generate, inputs=[instruction, system_prompt, version, temperature, max_new_tokens, top_p, repetition_penalty], outputs=[output, output2], ) demo.queue(concurrency_count=16).launch(debug=True)