schema_version int64 | snapshot_name string | snapshot_date timestamp[s] | archive dict | platform dict | images list | verification dict | upstream dict |
|---|---|---|---|---|---|---|---|
1 | olmes-code-executors-linux-amd64-snapshot-20261117 | 2026-11-17T00:00:00 | {
"filename": "olmes-code-executors-linux-amd64-snapshot-20261117.tar.zst",
"format": "docker-save-tar-zstd",
"compressed_size_bytes": 3941365258,
"uncompressed_archive_size_bytes": 3997891584,
"sha256": "ace98b640ce1b001175c373ffbd7803ef79431b5ecf5fec9b0ea8d604f60e969"
} | {
"os": "linux",
"architecture": "amd64"
} | [
{
"tag": "olmo-bcb-execution:latest",
"task_family": "BigCodeBench",
"image_id": "sha256:afe3f2fc8719f2720fd5754e72f0fd028488f6e076593afdf26941afbe84aeab",
"created": "2026-11-10T21:49:13.705803419+08:00",
"docker_image_ls_size": "6.64GB"
},
{
"tag": "olmo-ds1000-execution:latest",
"... | {
"credential_pattern_scan": "passed",
"zstd_integrity_test": "passed",
"docker_load_round_trip": "passed",
"loaded_image_ids_match_source": true
} | {
"olmes_repository": "https://github.com/allenai/olmes"
} |
OLMES code execution Docker images
This repository contains an unofficial, fixed linux/amd64 snapshot of the
three Docker execution environments used by the external-user code-evaluation
path in AllenAI OLMES:
| Image tag | OLMES task family | Expanded size shown by Docker |
|---|---|---|
olmo-bcb-execution:latest |
BigCodeBench | 6.64 GB |
olmo-ds1000-execution:latest |
DS-1000 | 6.00 GB |
olmo-multipl-execution:latest |
MultiPL-E HumanEval and MBPP | 4.33 GB |
The images were used successfully in OLMES evaluation runs. They are provided to make the execution environment easier to transfer to CPU executor nodes. This is an artifact snapshot, not an official AllenAI image release.
Download and load
Install the Hugging Face CLI and Zstandard, then run:
hf download Dawn9705/olmes-code-executor-images \
olmes-code-executors-linux-amd64-snapshot-20261117.tar.zst \
SHA256SUMS \
scripts/verify_images.sh \
--repo-type dataset \
--local-dir ./olmes-code-executor-images
cd ./olmes-code-executor-images
sha256sum -c SHA256SUMS
zstd -dc olmes-code-executors-linux-amd64-snapshot-20261117.tar.zst | docker load
bash scripts/verify_images.sh
Alternatively, scripts/download_and_load.sh performs all of these steps:
bash scripts/download_and_load.sh
The compressed archive is about 3.67 GiB. Allow roughly 17 GB of Docker image storage after loading it.
Snapshot identity
The archive keeps the image tags expected by OLMES. Do not use the mutable
latest label as the snapshot identity; use the full IDs below:
| Image | Full image ID | Image creation time |
|---|---|---|
| BigCodeBench | sha256:afe3f2fc8719f2720fd5754e72f0fd028488f6e076593afdf26941afbe84aeab |
2026-11-10 21:49:13 +08:00 |
| DS-1000 | sha256:aa113c7b9b51dcab02f51a8ce91623aca9794c21873f5e19f01fcbf79c467fda |
2026-11-10 21:53:38 +08:00 |
| MultiPL-E | sha256:8b7fd428da677ac9ca683968e83eeab6c0f8666bd044e499ec46ea9bb40b25e6 |
2026-08-15 23:33:53 +08:00 |
Machine-readable provenance is available in snapshot_manifest.json.
OLMES usage note
Only BigCodeBench, DS-1000, and MultiPL-E use these Docker images in the external-user OLMES route. HumanEval, MBPP, DeepSeek LeetCode, and Code FIM are executed by OLMES locally and do not use these images.
DS-1000 should be run with one Uvicorn worker in the OLMES remote-Docker setup. Using multiple server workers can trigger a Python multiprocessing pickling failure in its correctness checker.
Security
These containers execute model-generated, untrusted code. Use isolated, disposable executor hosts and do not expose the Docker daemon or executor HTTP ports to untrusted networks. The images must not be treated as a complete security boundary by themselves.
Before publication, the snapshot was checked for credential-like files and
patterns in its application directories, its image configuration was reviewed,
the Zstandard archive was tested, and a docker load round trip reproduced all
three original image IDs.
Licensing and provenance
This snapshot bundles many independently licensed components and therefore is
not distributed under one blanket license. See THIRD_PARTY.md. Users remain
responsible for complying with the licenses of the included upstream software.
- Downloads last month
- 5