autogen/samples/tools/testbed/utils/expand_gaia.py
afourney f8b4b4259b
Adds the GAIA benchark to the Testbed. This PR depends on #792 (#810)
* Re-added completion logging when using older versions of autogen.

* Extended scenario definitions and templating to include folders.

* Prepare collate_human_eval.py for working with group chat scenarios.

* Converted HumanEval to the folder-based approach, and added GroupChat scenarios.

* Fixed the default termination message.

* Fixed another termination condition.

* Updated compatible autogen versions.

* Added initial support for GAIA benchmark.

* Fixed a bug in executing the finalize scripts.

* Generalized the template further to support multiple folder copy operations.

* Refined GAIA support, and broke scenarios down by difficulty.

* Added some experimental scripts for computing metrics over GAIA. This is a first version, and will likely need refinement.

* Added instructions for cloning GAIA

* Updated README to fix some typos.

* Added a script to format GAIA reslts for the leaderboard.

* Update samples/tools/testbed/scenarios/GAIA/Templates/BasicTwoAgents/scenario.py

Co-authored-by: LeoLjl <3110503618@qq.com>

---------

Co-authored-by: Qingyun Wu <qingyun.wu@psu.edu>
Co-authored-by: LeoLjl <3110503618@qq.com>
2023-12-06 01:46:10 +00:00

111 lines
4.4 KiB
Python

#
# Run this file to download the human_eval dataset, and create a corresponding testbed scenario:
# (default: ../scenarios/human_eval_two_agents_gpt4.jsonl and ./scenarios/human_eval_two_agents_gpt35.jsonl)
#
import json
import os
import sys
import shutil
SCRIPT_PATH = os.path.realpath(__file__)
SCRIPT_NAME = os.path.basename(SCRIPT_PATH)
SCRIPT_DIR = os.path.dirname(SCRIPT_PATH)
SCENARIOS_DIR = os.path.realpath(os.path.join(SCRIPT_DIR, os.path.pardir, "scenarios", "GAIA"))
def create_jsonl(name, tasks, template, model):
"""Creates a JSONL scenario file with a given name, list of HumanEval tasks, template path, and model."""
with open(os.path.join(SCENARIOS_DIR, name + ".jsonl"), "wt") as fh:
for task in tasks:
print(f"Converting: [{name}] {task['task_id']}")
# Figure out what files we need to copy
template_cp_list = [template]
if len(task["file_name"].strip()) > 0:
template_cp_list.append(
[
os.path.join("GAIA_Files", task["file_name"].strip()),
os.path.join("coding", task["file_name"].strip()),
]
)
record = {
"id": task["task_id"],
"template": template_cp_list,
"substitutions": {
"scenario.py": {
"__MODEL__": model,
"__FILE_NAME__": task["file_name"],
"__PROMPT__": task["Question"],
},
"expected_answer.txt": {"__EXPECTED_ANSWER__": task["Final answer"]},
},
}
fh.write(json.dumps(record).strip() + "\n")
###############################################################################
if __name__ == "__main__":
if len(sys.argv) != 2:
sys.exit(
f"SYNTAX: python {SCRIPT_NAME} [path to GIA repository]\n\nNote: to clone the GAIA repository, do 'git clone https://huggingface.co/datasets/gaia-benchmark/GAIA'"
)
# Copy the relevant GAIA files
gaia_path = os.path.realpath(sys.argv[1])
gaia_validation_files = os.path.join(gaia_path, "2023", "validation")
gaia_test_files = os.path.join(gaia_path, "2023", "test")
if not os.path.isdir(gaia_validation_files) or not os.path.isdir(gaia_test_files):
sys.exit(f"Error: '{gaia_path}' does not appear to be a copy of the GAIA repository.")
gaia_merged_files = os.path.realpath(os.path.join(SCENARIOS_DIR, "GAIA_Files"))
shutil.copytree(
gaia_validation_files, gaia_merged_files, ignore=shutil.ignore_patterns("metadata.jsonl"), dirs_exist_ok=True
)
shutil.copytree(
gaia_test_files, gaia_merged_files, ignore=shutil.ignore_patterns("metadata.jsonl"), dirs_exist_ok=True
)
# Load the GAIA data
gaia_validation_tasks = [[], [], []]
with open(os.path.join(gaia_validation_files, "metadata.jsonl")) as fh:
for line in fh:
data = json.loads(line)
gaia_validation_tasks[data["Level"] - 1].append(data)
gaia_test_tasks = [[], [], []]
with open(os.path.join(gaia_test_files, "metadata.jsonl")) as fh:
for line in fh:
data = json.loads(line)
gaia_test_tasks[data["Level"] - 1].append(data)
models = {
"gpt4": "gpt-4",
}
templates = {
"two_agents": "Templates/BasicTwoAgents",
}
# Add coding directories if needed (these are usually empty and left out of the repo)
for template in templates.values():
code_dir_path = os.path.join(SCENARIOS_DIR, template, "coding")
if not os.path.isdir(code_dir_path):
os.mkdir(code_dir_path)
# Create the various combinations of [models] x [templates]
for m in models.items():
for t in templates.items():
create_jsonl(f"gaia_validation_level_1__{t[0]}_{m[0]}", gaia_validation_tasks[0], t[1], m[1])
create_jsonl(f"gaia_validation_level_2__{t[0]}_{m[0]}", gaia_validation_tasks[1], t[1], m[1])
create_jsonl(f"gaia_validation_level_3__{t[0]}_{m[0]}", gaia_validation_tasks[2], t[1], m[1])
create_jsonl(f"gaia_test_level_1__{t[0]}_{m[0]}", gaia_test_tasks[0], t[1], m[1])
create_jsonl(f"gaia_test_level_2__{t[0]}_{m[0]}", gaia_test_tasks[1], t[1], m[1])
create_jsonl(f"gaia_test_level_3__{t[0]}_{m[0]}", gaia_test_tasks[2], t[1], m[1])