Downloading Raw Results
MedHELM leaderboard results are stored in Google Cloud Storage (GCS) in the public crfm-helm-public bucket. Use the gcloud storage CLI (documentation) to download them.
Note: This page covers MedHELM results. Historical Stanford HELM projects (Classic, VHELM, AHELM, Image2Struct) may still exist under other paths in the same bucket but are not part of this repository.
Setup
- Install
gcloudfollowing Google’s instructions. Login is optional for the public bucket. - Create a local directory:
export LOCAL_BENCHMARK_OUTPUT_PATH=./benchmark_output mkdir -p $LOCAL_BENCHMARK_OUTPUT_PATH - Set the MedHELM GCS path:
export GCS_BENCHMARK_OUTPUT_PATH=gs://crfm-helm-public/medhelm/benchmark_output
MedHELM path
- MedHELM:
gs://crfm-helm-public/medhelm/benchmark_output
Download the full MedHELM tree
Warning: the full tree can be large. Check size before downloading:
gcloud storage du -sh $GCS_BENCHMARK_OUTPUT_PATH
gcloud storage rsync -r $GCS_BENCHMARK_OUTPUT_PATH $LOCAL_BENCHMARK_OUTPUT_PATH
Download a specific release
- Set the release version (see leaderboard.medhelm.org for available versions):
export RELEASE_VERSION=v2.0.0 mkdir -p $LOCAL_BENCHMARK_OUTPUT_PATH/releases/$RELEASE_VERSION gcloud storage rsync -r \ $GCS_BENCHMARK_OUTPUT_PATH/releases/$RELEASE_VERSION \ $LOCAL_BENCHMARK_OUTPUT_PATH/releases/$RELEASE_VERSION - Inspect
$LOCAL_BENCHMARK_OUTPUT_PATH/releases/$RELEASE_VERSION/summary.jsonfor suite names, then download individual suites if needed (below).
Download a specific suite
export SUITE_VERSION=<suite_name_from_summary.json>
mkdir -p $LOCAL_BENCHMARK_OUTPUT_PATH/runs/$SUITE_VERSION
gcloud storage rsync -r \
$GCS_BENCHMARK_OUTPUT_PATH/runs/$SUITE_VERSION \
$LOCAL_BENCHMARK_OUTPUT_PATH/runs/$SUITE_VERSION
Troubleshooting
On older gcloud versions, du or rsync subcommands may be missing. Upgrade gcloud, or use gsutil du / gsutil rsync instead (gsutil docs).
GCS browser
Browse files in the GCS console. This requires a Google account and acceptance of GCP Terms of Service.
For a smaller download when you only need aggregated metrics, see LEADERBOARD_EXPORT.md.