import os import sys import gradio as gr from dataclasses import dataclass from chalk import * from colour import Color import inspect import os import openai from typing import List, Tuple, Optional from enum import Enum import io from contextlib import redirect_stdout import imageio import tiktoken import time import pandas as pd import csv from huggingface_hub import HfApi, Repository DATASET_REPO_URL = "https://huggingface.co/datasets/srush/gptworld-leaderboard" HF_TOKEN = os.environ.get("HF_API") DATA_FILENAME = "data.csv" DATA_FILE = os.path.join("data", DATA_FILENAME) openai.api_key = "" tab = " " repo = Repository(git_user="srush", local_dir="data", clone_from=DATASET_REPO_URL, use_auth_token=HF_TOKEN ) repo.git_pull() def start2(prompt, board, api_key): out = "" # for chunk in openai.ChatCompletion.create( # model="gpt-4", # messages=[{ # "role": "user", # "content": prompt, # }], # stream=True, # temperature= 0 # ): board = board#Game(boundary=(9, 9), key=(1, 1), flag=(2, 2), init=(0, 0), walls=[(2, 0)]) actions = [Actions.DOWNRIGHT, Actions.RIGHT, Actions.DOWNRIGHT, Actions.PICKUP, Actions.DOWNRIGHT] contents = example(board, actions) print(contents) # encoding = tiktoken.encoding_for_model("gpt-4") # num_tokens = encoding.encode(string) for content in contents: time.sleep(0.005) content = content if content is not None: out += content print(content, end="") yield out yield out def start(prompt, board, api_key): out = "" # encoding = tiktoken.encoding_for_model("gpt-4") # num_tokens = encoding.encode(string) content = "" openai.api_key = api_key for chunk in openai.ChatCompletion.create( model="gpt-4", messages=[{ "role": "user", "content": prompt, }], stream=True, temperature= 0 ): # for content in contents: time.sleep(0.005) content = chunk["choices"][0].get("delta", {}).get("content") if content is not None: out += content print(content, end="") yield out yield out def num_tokens_from_string(string: str, encoding_name: str="gpt-4") -> int: """Returns the number of tokens in a text string.""" encoding = tiktoken.encoding_for_model(encoding_name) num_tokens = len(encoding.encode(string)) return num_tokens # + [markdown] id="LMTjwXdD7v-I" # ## Game Code # # This code creates a mini-game to play. It takes place on a hexagon. You are represented by a circle. You need to first pick up a key represented by a triangle. You finally need to make it to the cross to finish the game. The actions show each of the directions you can move. # # # + id="Fv3eTRKiV2ZB" cellView="form" #@title Game Code # Possible Actions class Actions(Enum): UPRIGHT = "UR" RIGHT = "R" DOWNRIGHT = "DR" DOWNLEFT = "DL" LEFT = "L" UPLEFT = "UL" PICKUP = "Pickup" # Movements change = { Actions.UPRIGHT : (-1, 1), Actions.RIGHT : (0, 2), Actions.DOWNRIGHT : (1, 1), Actions.DOWNLEFT : (1, -1), Actions.LEFT : (0, -2), Actions.UPLEFT : (-1, -1), Actions.PICKUP : (0, 0), } change_str = {action.value: change[action] for action in Actions} def add(a, b): return a[0] + b[0], a[1] + b[1] @dataclass class Board: grid: List[str] player_pos: Tuple[int, int] flag_pos: Tuple[int, int] wall_pos:List[Tuple[int, int]] key_pos:Optional[Tuple[int, int]] def move(self, action: Actions) -> 'Board': "Move by creating a new board." d_m = change[action] if action == Actions.PICKUP: if self.player_pos == self.key_pos: return Board(self.grid, self.player_pos, self.flag_pos, self.wall_pos, None) else: return self new_player_pos = add(self.player_pos, d_m) # Out of bounds if new_player_pos[0] < 0 or new_player_pos[0] >= len(self.grid): return self if new_player_pos[1] < 0 or new_player_pos[1] >= len(self.grid[0]): return self # Can't move through walls if self.grid[new_player_pos[0]][new_player_pos[1]] == 'W': return self new_grid = [row[:] for row in self.grid] # Create a copy of the grid new_grid[self.player_pos[0]][self.player_pos[1]] = '.' new_grid[new_player_pos[0]][new_player_pos[1]] = '@' return Board(new_grid, new_player_pos, self.flag_pos, self.wall_pos, self.key_pos) def __str__(self) -> str: return '\n'.join(''.join(row) for i, row in enumerate(self.grid)) @classmethod def create_empty_board(cls, size: Tuple[int, int], key_pos, flag_pos, init, wall_pos) -> 'Board': grid = [['.' if i % 2 == j % 2 else " " for i in range(size[1])] for j in range(size[0])] player_pos = init flag_pos = flag_pos grid[player_pos[0]][player_pos[1]] = '@' grid[flag_pos[0]][flag_pos[1]] = 'P' grid[key_pos[0]][key_pos[1]] = 'K' for pos in wall_pos: grid[pos[0]][pos[1]] = 'W' return cls(grid, player_pos, flag_pos, wall_pos, key_pos) class Game: def __init__(self, init, flag, walls, key, boundary): "Create the version of the game that the AI sees." self.boundary = boundary self.board = Board.create_empty_board(boundary, key, flag, init, walls) self.original = self.board self.actions = [] def move(self, action): self.board = self.board.move(action) self.actions.append(action) @property def walls(self): return self.board.wall_pos def won(self): final = self.board return final.key_pos is None and final.player_pos == final.flag_pos def __repr__(self) -> str: walls = ",".join(map(str, self.board.wall_pos)) return f"Game(init={self.board.player_pos}, flag={self.board.flag_pos}, walls={self.board.wall_pos}, boundary={self.boundary}, key={self.board.key_pos})" # This is the version of move that the AI can see. def move(game, action, old_pos): # ACTIONS (must be legal) game.move(Actions(action)) offset = change_str[action] pos = (old_pos[0] + offset[0], old_pos[1] + offset[1]) assert 0 <= pos[0] < game.boundary[0], "Row position out of bounds" assert 0 <= pos[1] < game.boundary[1], "Col position out of bounds" assert pos not in game.walls, f"Walked into wall {pos}" if action == "PU": assert pos == game.key, f"Not over key" return pos # + [markdown] id="PDOcPiQq8u_Y" # We can look at the board by drawing it. # + colab={"base_uri": "https://localhost:8080/", "height": 221} id="Ic7WgOTi8uF1" outputId="4dc07cb9-9e5f-4d28-d4ea-470ad4b13141" #@title Drawing code def draw_board(grid, num=0): hex = regular_polygon(6, 1).rotate_by(1/12).line_width(0.5).fill_color(Color("white")) w = hex.get_envelope().width canvas = empty() for r, b in enumerate(grid): def show(v): if v == ".": return hex if v == "@": return hex + circle(0.35).fill_color(Color("red")) if v == "P": x = rectangle(0.25, 0.7).fill_color(Color("blue")).line_width(0) return hex + (x.rotate_by(0.25/2) + x.rotate_by(-0.25/2)) if v == "K": return hex + triangle(0.75).fill_color(Color("purple")) if v == "W": return hex.fill_color(Color("black")) if v ==" ": return hex row = hcat(show(v) for i, v in enumerate(b[1 if r %2 else 0::2])) canvas += row.translate(w * 0.5 if r%2 else 0, 1.5 * r) canvas = canvas.center_xy().frame(0.5) canvas = rectangle(canvas.get_envelope().width, canvas.get_envelope().height).line_width(0.5).fill_color(Color("orange")) + canvas canvas.render_svg(f"pic{num}.svg", 256) canvas.render(f"pic{num}.png", 256) return canvas # + colab={"base_uri": "https://localhost:8080/", "height": 424} id="nqgPKLu0AMhU" outputId="19e4c6d0-b792-4a34-f4c4-81902974c346" # game = Game(boundary=(5, 5), key=(0, 2), flag=(4, 4), init=(0, 0), walls=[(2, 2)]) # display(draw_board(game.board.grid)) # move(game, "DR", (0,0)) # display(draw_board(game.board.grid)) # + [markdown] id="PhqF9af5_jvh" # ## Prompt Code # # The puzzle is to write prompt code to make the model accomplish this task. We have provided some scaffolding code for you. The code creates: # # * A header for describing the game. # * A function `make_fun` that shows the AI how to move in code. # * A footer to describe the final game board that you want the mode to solve. # # You can fill this in a watch how the model moves around. # + id="jFf7TCOJaVHX" #@title Make the Prompt def make_fun(board, actions): "This function generates python code for few-shot examples" out = tab + "p = " + str(board.player_pos) for i, action in enumerate(actions): new_board = board.move(action) out += f""" # TODO ADD CODE p = move(b, "{action.value}", p) # TODO ADD CODE""" board = new_board return out def example(game, actions): """ This code makes a few shot example. You don't need to edit it. """ return f""" def my_example(): b = {repr(game)} {make_fun(game.board, actions)} return b """ ex = 0 def prompt(game): """ You should fill these sections out to teach the AI how to play the game. Or you may do your own thing :) """ print(f""" # TODO: DESCRIBE THE GAME # TODO: DESCRIBE THE ACTIONS change_str = {change_str} {inspect.getsource(move)} """) def example(game, actions): """ This code makes a few shot example. You don't need to edit it. """ global ex ex += 1 print(f""" def example{ex}(): b = {repr(game)} {make_fun(game.board, actions)} return b # ------------ """) # Create a few shot example (you may not need this) board = Game(boundary=(3, 3), key=(1, 1), flag=(2, 2), init=(0, 0), walls=[(2, 0)]) actions = [Actions.DOWNRIGHT, Actions.PICKUP, Actions.DOWNRIGHT] example(board, actions) # Test case print(f""" # ---- # TODO: ADD any custom example code #--- # TODO: FINAL description. # Contraints for this function:", {repr(game)} # Please fill this in with code like the examples above (do not provide a description): # # The following function `my_example` instantiates a GameBoard called b with these constraints. """) # + [markdown] id="-iecyV7nAbFT" # This code lets you make a game and see the output for a prompt for that game. There are easy, medium, and hard games. # + colab={"base_uri": "https://localhost:8080/"} id="cOneYFok_OMe" outputId="97080186-7322-4ba9-b500-095fb39071aa" # Easy easy_game = Game(boundary=(3, 3), key=(1, 1), flag=(2, 2), init=(0, 0), walls=[]) # Medium medium_game = Game(boundary=(5, 5), key=(3, 1), flag=(4, 4), init=(0, 0), walls=[(1, 1)]) # Hard (This is the main one) hard_game = Game(boundary=(8, 15), key=(3, 1), flag=(7, 13), init=(0, 0), walls=[(2, 2), (1, 1), (5, 3), (1, 11), (5, 5), (6, 6), (6, 10), (2, 6), (4, 12)]) # Evil evil_game = Game(boundary=(8, 15), key=(5, 1), flag=(7, 13), init=(0, 0), walls=[(2, 2), (3, 3), (4, 2), (1, 1), (2, 4), (7, 11), (5, 3), (1, 11), (5, 5), (6, 6), (6, 10), (2, 6), (4, 12)]) games = {"Easy": easy_game, "Medium": medium_game, "Hard": hard_game, "Evil": evil_game} # Anima def animate(game): cur = game.original i = 0 images = [] draw_board(cur.grid, i) images.append(imageio.v2.imread(f"pic{i}.png")) for act in game.actions: cur = cur.move(act) i += 1 draw_board(cur.grid, i) images.append(imageio.v2.imread(f"pic{i}.png")) return imageio.v2.mimsave('movie.gif', images, **{ 'duration': 1000, 'loop': 100}) def load(inp): if inp in games: board = games[inp] else: board = eval(inp) draw_board(board.board.grid, 0).render_svg("tmp.svg") return ["tmp.svg"], repr(board) draw_board(hard_game.board.grid, 0).render_svg("hard.svg") draw_board(easy_game.board.grid, 0).render_svg("easy.svg") with gr.Blocks() as app: # test = gr.Code(label="test") # im2 = gr.Gallery() # im2.style(preview=True) gr.HTML("""

🌎 GPTWorld 🌍

GPTWorld is a prompting game. Your goal is to get an LLM to complete a maze. You are the red dot (πŸ”΄) need to first get the key (β–²) and then reach the exit (x). The game takes place on a hexagonal grid with walls [Even rows are labeled (0,0), (0, 2), (0,4) and odd rows are labeled (1, 1), (1, 3), (1, 5)]. You play by prompting GPT to write code which solves the game on the right.

""") with gr.Row(): with gr.Column(): game_desc = gr.Text(label="Game (Select one first or make your own)") examples = gr.Radio(show_label=False, choices=["Easy", "Medium", "Hard", "Evil"]) api_key = gr.Text(label="OpenAI Key", type="password", value=os.environ.get("OPENAI_API_KEY"), visible=not os.environ.get("OPENAI_API_KEY")) with gr.Row(): start_btn = gr.Button("Prompt >") cancel_btn = gr.Button("Cancel") prompt = gr.Code(label="Prompt (This is for you to fill out)", language="python", lines=40, value=f""" # A prompt to describe this game to the GPT model. # Ideas: # * Describe how the game works # * Give code examples that solve similar mazes. # * Give examples to explain the reasoning process # You might want to do this in a separate editor and paset in. # For example you might want to tell it how the moves work change_str = {change_str} # Or make up a clear implementation for the move function. def move(board, action, old_pos): # ACTIONS (must be legal) board.move(action) offset = change_str[action] pos = (old_pos[0] + offset[0], old_pos[1] + offset[1]) assert 0 <= pos[0] < board.boundary[0] assert 0 <= pos[1] < board.boundary[1] assert pos not in board.walls if action == "Pickup": assert pos == board.key return pos # You can also test your code on the right side to make few-shot examples. # Finally use %GAME% to inject the game description above. """) with gr.Column(): im = gr.Gallery(label="Gallery of the Game") im.style(preview=True, object_fit="scale-down", columns=1, container=True) msg_box = gr.HTML(label="", show_label=False) output = gr.Code(label="Generating Game Code (You can also edit and rerun)", language="python", value="""def my_example(): b = Game(init=(0, 0), flag=(2, 2), walls= [], boundary= (3, 3), key= (1, 1)) p = (0, 0) # This is the code you want it to generate. p = move(b, "DR", p) p = move(b, "Pickup", p) p = move(b, "DL", p) p = move(b, "R", p) return b """, lines=50) counter = gr.Slider(label="length", minimum=0, maximum=3000) run_btn = gr.Button("Rerun ^") state = gr.State() examples.change(load, inputs=[examples], outputs=[im, game_desc]) game_desc.submit(load, inputs=[game_desc], outputs=[im, game_desc]) def run(data): board = eval(data[game_desc]) #games[data[examples]] inp = data[prompt].replace("%GAME%", repr(board)) print(inp) q = {} i = 0 count = 0 im_ = [f"tmp.svg"] state_val = None yield {im: im_, counter: 0, output: "", msg_box: "", state: state_val} for prefix in start(inp, board, data[api_key]): ps = prefix.split("\n") count += 1 if len(ps) > 3 and not ps[-2].strip().startswith("#") and prefix.endswith("\n"): print("rendering") try: exec(prefix + f"\n return b\nq['board'] = my_example({repr(board)})") except AssertionError as e: print("fail") yield {im: [f"pic{j}.svg" for j in range(i)], counter: count, output: prefix, msg_box: f"You made an illegal move: {e}"} return draw_board(q["board"].board.grid, i).render_svg("tmp.svg") i += 1 im_ = [f"pic{j}.svg" for j in [i-1]] yield {im: im_, counter: count, output: prefix} else: yield {im: im_, counter: count, output: prefix} if q["board"].won(): final_msg = "πŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡VictoryπŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡πŸŽ‡" state_val = (data[prompt], prefix, count, data[examples]) else: final_msg = "Didn't make it" animate(q["board"]) yield {im: ["movie.gif"], counter: count, output: prefix, msg_box: final_msg, state: state_val} start_prompt = start_btn.click(run, inputs={prompt, game_desc, api_key, examples}, outputs={im, output, counter, msg_box, state}) cancel_btn.click(None, cancels=[start_prompt]) def run2(data): c = data[output] print(c) i = 0 q = {} for j in range(len(c)): prefix = c[:j] ps = prefix.split("\n") if len(ps) > 3 and not ps[-2].strip().startswith("#") and prefix.endswith("\n"): print("rendering", prefix) exec(prefix + "\n return b\nq['board'] = my_example()") draw_board(q["board"].board.grid, i) i += 1 animate(q["board"]) out = {im: [f"movie.gif"], msg_box: ""} print(out) return out run_btn.click(run2, inputs={output}, outputs={im, msg_box}) gr.HTML("""

Leaderboard

To submit, first run a model that gets to the end. It will output "Victory"! Then come down to the bottom, type in your team name, and then click submit. The score is the number of output tokens your solution takes.
""") with gr.Row() as row: team_name = gr.Text(label="Team Name") leaderboard = gr.Button(value="Submit") msg = gr.Text(label="Status") leader = gr.Dataframe(pd.read_csv(DATA_FILE)[["team", "board", "count"]]) def leaderfn(data): if data[state] is None: return {msg: "Nothing to submit"} if not data[team_name]: return {msg: "No team name"} prompt, code, count, board = data[state] repo.git_pull() with open(DATA_FILE, "a") as csvfile: writer = csv.DictWriter(csvfile, fieldnames=["team", "prompt", "code", "count", "board"]) writer.writerow( {"team": data[team_name], "prompt": prompt, "code": code, "count": count, "board": board} ) commit_url = repo.push_to_hub() leader_df = pd.read_csv(DATA_FILE)[["team", "board", "count"]] leader_df = leader_df.sort_values(["board", "count"]) return {msg: f"Sucess: Final score: {count} {board}", leader: leader_df} leaderboard.click(fn=leaderfn, inputs={state, team_name}, outputs={msg, leader}) app.queue().launch() # f = io.StringIO() # with redirect_stdout(f): # ex = 0 # prompt(game) # my_prompt = f.getvalue() # print(my_prompt) # # + id="LONWUsBLjOHo" colab={"base_uri": "https://localhost:8080/", "height": 1000} outputId="472afd19-48c1-4924-cabd-639b5e2ad298" # # Run an LLM and execute it as it runs. # q = {} # i = 0 # for prefix in start(my_prompt): # ps = prefix.split("\n") # if len(ps) > 3 and not ps[-2].strip().startswith("#") and prefix.endswith("\n"): # exec(prefix + "\n return b\nq['board'] = my_example()") # display(draw_board(q["board"].board.grid, i)) # i += 1 # animate(i) # display(Image("movie.gif")) # # Print the number of tokens used # print("Input Tokens:", num_tokens_from_string(my_prompt)) # print("Output Tokens:", num_tokens_from_string(prefix))