Release 0.3.0 (ICML version, only L1) (#17)

* Internal repo sync

* version bump 0.3.0

* Update README.md

---------

Co-authored-by: Leo Boisvert <[email protected]>
Co-authored-by: Alexandre Drouin <[email protected]>

Author: 3 people

README.md

@@ -10,12 +10,9 @@ WorkArena is included in [BrowserGym](https://github.com/ServiceNow/BrowserGym),
 
 https://github.com/ServiceNow/WorkArena/assets/2374980/68640f09-7d6f-4eb1-b556-c294a6afef70
 
-## ⚠️ Pre-Release warning ⚠️
-Please note that the WorkArena benchmark is still undergoing minor bug fixes and updates, which may cause discrepancies with results reported in our latest arXiv preprint. We plan to release soon a stable version of WorkArena with enhanced stability, and a final version v1.0.0 with a new suite of tasks.
-
 ## Benchmark Contents
 
-At the moment, WorkArena includes `19,951` task instances drawn from `33` tasks that cover the main components of the ServiceNow user interface. The following videos show an agent built on `GPT-4-vision` interacting with every such component. As emphasized by our results, this benchmark is not solved and thus, the performance of the agent is not always on point.
+At the moment, WorkArena includes `19,912` unique instances drawn from `33` tasks that cover the main components of the ServiceNow user interface. The following videos show an agent built on `GPT-4-vision` interacting with every such component. As emphasized by our results, this benchmark is not solved and thus, the performance of the agent is not always on point.
 
 ### Knowledge Bases
 
@@ -53,8 +50,11 @@ https://github.com/ServiceNow/WorkArena/assets/1726818/ca26dfaf-2358-4418-855f-8
 
 ### Dashboards
 
-**Goal:** The agent must extract information from a dashboard.
+**Goal:** The agent must answer a question that requires reading charts and (optionally) performing simple reasoning over them.
+
+*Note: For demonstration purposes, a human is controlling the cursor since this is a pure retrieval task*
 
+https://github.com/ServiceNow/WorkArena/assets/1726818/0023232c-081f-4be4-99bd-f60c766e6c3f
 
 
 ## Getting Started
@@ -98,6 +98,8 @@ Your installation is now complete! 🎉
 
 Run this code to see WorkArena in action.
 
+Note: the following example executes WorkArena's oracle (cheat) function to solve each task. To evaluate an agent, calls to `env.step()` must be used instead.
+
 ```python
 import random
 
@@ -112,28 +114,27 @@ for task in ALL_WORKARENA_TASKS:
 
     # Instantiate a new environment
     env = BrowserEnv(task_entrypoint=task,
-                    headless=False, 
-                    slow_mo=1000)
+                    headless=False)
     env.reset()
 
     # Cheat functions use Playwright to automatically solve the task
     env.chat.add_message(role="assistant", msg="On it. Please wait...")
-    env.task.cheat(env.page, env.chat.messages)
+    cheat_messages = []
+    env.task.cheat(env.page, cheat_messages)
+
+    # Send cheat messages to chat
+    for cheat_msg in cheat_messages:
+        env.chat.add_message(role=cheat_msg["role"], msg=cheat_msg["message"])
 
     # Post solution to chat
-    if "KnowledgeBaseSearchTask" in str(task):
-        answer = env.chat.messages[-1]["message"]
-        env.chat.add_message(role="assistant", msg=f"The answer is:")
-        env.chat.add_message(role="assistant", msg=answer)
-    else:
-        env.chat.add_message(role="assistant", msg="I'm done!")
+    env.chat.add_message(role="assistant", msg="I'm done!")
 
     # Validate the solution
-    reward, stop, info, message = env.task.validate(env.page, env.chat.messages)
+    reward, stop, message, info = env.task.validate(env.page, cheat_messages)
     if reward == 1:
         env.chat.add_message(role="user", msg="Yes, that works. Thanks!")
     else:
-        env.chat.add_message(role="user", msg=f"No, that doesn't work. {message.get('message', '')}")
+        env.chat.add_message(role="user", msg=f"No, that doesn't work. {info.get('message', '')}")
 
     sleep(3)
     env.close()

pyproject.toml

@@ -31,6 +31,7 @@ homepage = "https://github.com/ServiceNow/WorkArena"
 
 [project.scripts]
 workarena-install = "browsergym.workarena.install:main"
+workarena-human-eval = "browsergym.workarena.human_eval.tool:main"
 
 [tool.hatch.version]
 path = "src/browsergym/workarena/__init__.py"

requirements.txt

@@ -1,6 +1,6 @@
 browsergym-core>=0.2
 english-words>=2.0.1
-faker>=24.11.0
+Faker>=24.8.0
 numpy>=1.14
 requests>=2.31
 tenacity>=8.2.3  # only used in cheat() -> move to tests?

scripts/extract_finetuning_traces.py

@@ -0,0 +1,131 @@
+"""
+A demonstration of how observation/action traces can be extracted
+for WorkArena tasks without modifying the task code.
+
+Author: Alexandre Drouin ([email protected])
+
+Notes:
+- This approach relies on monkey patching the playwright actions to log the actions and observations.
+  It has not been tested for parallel execution. It might work with multiprocessing, but it will for
+  sure not work with multithreading.
+
+"""
+
+import importlib
+import logging
+import os
+import pickle
+import playwright.sync_api as playwright_sync
+
+from browsergym.core.env import BrowserEnv
+from browsergym.workarena import ALL_WORKARENA_TASKS
+from collections import defaultdict
+from tenacity import retry, stop_after_attempt, wait_fixed
+from time import time
+
+
+N_PER_TASK = 10
+
+
+def monkey_patch_playwright(observation_callback, trace_storage):
+    """
+    A function that overrides the default playwright actions to log the actions and observations.
+
+    Parameters:
+    ------------
+    observation_callback: callable
+        A function that returns the observation of the environment.
+    trace_storage: list
+        A list to store the trace of the actions and observations.
+        These will be appended in-place.
+
+    """
+
+    def wrapper(func, interface):
+        def wrapped(*args, **kwargs):
+            # Get the observation
+            obs = observation_callback()
+
+            # Get the BID of the element on which we are acting.
+            if interface.__name__ == "Locator":
+                # Get the locator
+                locator = args[0]
+                # Get the BID
+                bid = locator.element_handle().evaluate('(el) => el.getAttribute("bid")')
+            elif interface.__name__ == "Keyboard":
+                # Get the BID of the element
+                bid = "keyboard"
+            else:
+                # Get the BID of the element
+                bid = args[0].evaluate('(el) => el.getAttribute("bid")')
+
+            logging.info(f"Action: {func.__name__} BID: {bid}  --   Args: {args[1:]} {kwargs}")
+            trace_storage.append(
+                {
+                    "obs": obs,
+                    "action": func.__name__,
+                    "args": args[1:],
+                    "kwargs": kwargs,
+                    "bid": bid,
+                    "time": time(),
+                }
+            )
+
+            # Resume action
+            return func(*args, **kwargs)
+
+        return wrapped
+
+    # Interfaces and actions we want to monkey patch
+    importlib.reload(playwright_sync)
+    from playwright.sync_api import Page, Frame, Locator, Keyboard, ElementHandle
+
+    # TODO: Make sure the list of interfaces and actions is exhaustive
+    #       It covers all that is used in WorkArena cheats as of April 11, 2024
+    interfaces = [Page, Frame, Locator, Keyboard, ElementHandle]
+    actions = ["click", "select_option", "set_checked", "fill", "press", "type", "down", "up"]
+
+    for interface in interfaces:
+        for action in actions:
+            if hasattr(interface, action):
+                setattr(interface, action, wrapper(getattr(interface, action), interface))
+                print(f"Monkey patched {interface.__name__}.{action}")
+
+
+@retry(stop=stop_after_attempt(3), wait=wait_fixed(2))
+def extract_trace(task_cls, headless=True):
+    """
+    Extracts the trace of actions and observations for a given task.
+
+    Parameters:
+    ------------
+    task_cls: class
+        The class of the task to extract the trace from.
+
+    """
+    # Instantiate a new environment
+    env = BrowserEnv(task_entrypoint=task_cls, headless=headless, slow_mo=1000)
+
+    # Setup customized tracing
+    trace = []
+    monkey_patch_playwright(observation_callback=env._get_obs, trace_storage=trace)
+
+    env.reset()
+    env.task.cheat(env.page, env.chat.messages)
+    env.close()
+
+    return trace
+
+
+if __name__ == "__main__":
+    os.makedirs("trace_profiling", exist_ok=True)
+
+    task_traces = defaultdict(list)
+    for task in ALL_WORKARENA_TASKS:
+        print("Task:", task)
+        for i in range(N_PER_TASK):
+            print(f"Extracting trace {i+1}/{N_PER_TASK}")
+            trace = extract_trace(task, headless=True)
+            task_traces[task].append(trace)
+
+    pickle.dump(task_traces, open("trace_profiling/task_traces.pkl", "wb"))