Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add bsevalharness compat #5

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add bsevalharness compat #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions tasks/eval_harness/download_bsevalharness.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
# Downloads the specified taks in the evaluation harness
# This is particularly useful when running in environments where the GPU nodes
# do not have internet access. This way we can pre-download them and use the cached data-set during evaluation.

from lm_eval import tasks
from lm_eval.tasks import ALL_TASKS
import argparse
import os


parser = argparse.ArgumentParser(description='Download evaluation harness', allow_abbrev=False)
parser.add_argument('--task_list', type=str, default = "all", help='Either "all" or comma separated list of tasks to download.')
args = parser.parse_args()

def main():
task_list = ALL_TASKS if args.task_list == 'all' else args.task_list.split(',')
tasks.get_task_dict_promptsource(task_list)

if __name__ == '__main__':
main()
9 changes: 5 additions & 4 deletions tasks/eval_harness/evaluate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -399,6 +399,7 @@ def tasks_args(parser):
group.add_argument('--bootstrap_iters', type=int, default=100000, help='How many iterations to use for stderr estimation')
group.add_argument('--micro_bs_multiplier', type=int, default=1, help='Increase the global batch size to remove bubble when pipeline parallel')
group.add_argument('--add_denoiser', default = False, action='store_true', help='Whether to add a denoiser to the model')
group.add_argument('--fewshots', type=int, default=0, help='Number of fewshots')
return parser

from megatron.global_vars import _parse_args
Expand Down Expand Up @@ -431,11 +432,11 @@ def main():
global_results = {"results": {}, "versions": {}}
timestamp = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')
iteration_id = load_path.split("/")[-1].replace("/", "")
results_path = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}.json")
results_path = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}_{args.fewshots}shots.json")
# Backup file in case of interruption during writing
results_path_backup = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}_backup.json")
results_path_backup = args.results_path.replace(".json", f"_lm-eval_{iteration_id}_{timestamp}_{args.fewshots}shots_backup.json")
for task_name, task in task_dict.items():
results = evaluator.evaluate(adaptor, {task_name: task}, False, 0, None, bootstrap_iters=args.bootstrap_iters)
results = evaluator.evaluate(adaptor, {task_name: task}, False, args.fewshots, None, bootstrap_iters=args.bootstrap_iters)
global_results["results"] = {**global_results["results"], **results["results"]}
global_results["versions"] = {**global_results["versions"], **results["versions"]}
if mpu.is_pipeline_last_stage() and mpu.get_tensor_model_parallel_rank() == 0:
Expand All @@ -445,7 +446,7 @@ def main():
with open(results_path_backup, 'w') as outfile:
json.dump(global_results, outfile, indent=4)
else:
global_results = evaluator.evaluate(adaptor, task_dict, False, 0, None, bootstrap_iters=args.bootstrap_iters)
global_results = evaluator.evaluate(adaptor, task_dict, False, args.fewshots, None, bootstrap_iters=args.bootstrap_iters)
if mpu.is_pipeline_last_stage() and mpu.get_tensor_model_parallel_rank() == 0:
print(json.dumps(global_results, indent=2))
with open(args.results_path, 'w') as outfile:
Expand Down
Loading