OpenEvo scientific runtime — Docker image (exact evaluation artifact)
中文说明见下方「中文说明」一节。This card is bilingual.
This repository hosts the exact Docker runtime image used to run the SEED
official held-out WebShop comparison (20260825-0624-webshop-seed-official-heldout-comparison)
and to evaluate the frozen OpenEvo Qwen2.5-7B SD-LoRA policy. It is published so
independent agents can obtain the bit-identical runtime, not just a recipe.
The image is a docker save tarball. Download, load, and verify the digest before use.
Identity (verify before use)
| Item | Value |
|---|---|
| Image ref | wangr/openevo-scientific:qwen25-3b-webshop-20260814 |
| Image ID / digest | sha256:ce92e587e7c934827025c001fd9b70ec4cb6de2deacfdf9f0cc0777d3ab928f1 |
| Tarball file | openevo-scientific-qwen25-3b-webshop-20260814.tar.gz |
| Tarball size | 5,424,572,660 bytes (~5.4 GB, docker save | gzip -1) |
| Tarball SHA-256 | 8a2e1c4c2cf413bb590f5ba066a1eebbfa460675e58fe9a7b4fc87e500a0ec47 |
Load and verify
# 1. download the tarball from this repo, then verify its SHA-256
sha256sum openevo-scientific-qwen25-3b-webshop-20260814.tar.gz
# -> 8a2e1c4c2cf413bb590f5ba066a1eebbfa460675e58fe9a7b4fc87e500a0ec47
# 2. load into Docker
gunzip -c openevo-scientific-qwen25-3b-webshop-20260814.tar.gz | docker load
# 3. confirm the loaded image ID matches the pinned digest
docker image inspect wangr/openevo-scientific:qwen25-3b-webshop-20260814 --format '{{.Id}}'
# -> sha256:ce92e587e7c934827025c001fd9b70ec4cb6de2deacfdf9f0cc0777d3ab928f1
Reproducible build recipe (alternative to the tarball)
The image is built from a committed, pinned recipe in the evaluation repository
(mykcs/openevo-experiment): scripts/openevo_webshop/seed_runtime/Dockerfile
plus its lock file and full build context (PyTorch 2.13.0+cu130, vLLM 0.27.1,
FlashInfer 0.6.16.post4, Blackwell-compatible). If you prefer to build rather
than download 5.4 GB, use that recipe; the tarball here is the exact artifact the
evaluation pinned.
What runs inside
- Scientific Python and model processes run as non-root numeric UID/GID
1001:1001. - This is the runtime only. The evaluated model adapter is hosted separately at
miyuki17/openevo-qwen25-7b-webshop-sd-lora. The WebShop task data and SEED checkout are not in this image.
Claim boundary
This image is a reproducibility artifact. Publishing it makes no scientific claim;
all results and their validity boundaries are documented in the evaluation
repository under docs/evidence/seed-official-heldout-comparison-v1/.
中文说明
本仓库托管运行 SEED 官方 held-out WebShop 对比
（20260825-0624-webshop-seed-official-heldout-comparison）以及评测冻结 OpenEvo
Qwen2.5-7B SD-LoRA 政策时所用的同一个 Docker 运行时镜像。发布它是为了让独立
Agent 能拿到逐比特一致的运行时，而不仅是一份构建配方。
镜像以 docker save tarball 形式提供。使用前请先下载、docker load 并核对 digest。
身份（使用前必校验）
| 项 | 值 |
|---|---|
| 镜像引用 | wangr/openevo-scientific:qwen25-3b-webshop-20260814 |
| 镜像 ID / digest | sha256:ce92e587e7c934827025c001fd9b70ec4cb6de2deacfdf9f0cc0777d3ab928f1 |
| tarball 文件 | openevo-scientific-qwen25-3b-webshop-20260814.tar.gz |
| tarball 大小 | 5,424,572,660 字节（约 5.4 GB） |
| tarball SHA-256 | 8a2e1c4c2cf413bb590f5ba066a1eebbfa460675e58fe9a7b4fc87e500a0ec47 |
加载与校验
sha256sum openevo-scientific-qwen25-3b-webshop-20260814.tar.gz # 应为 8a2e1c4c…
gunzip -c openevo-scientific-qwen25-3b-webshop-20260814.tar.gz | docker load
docker image inspect wangr/openevo-scientific:qwen25-3b-webshop-20260814 --format '{{.Id}}' # 应为 sha256:ce92e587…
构建配方（可替代直接下载）
镜像由评测仓库（mykcs/openevo-experiment）中已提交、已锁定的配方构建：
scripts/openevo_webshop/seed_runtime/Dockerfile 及其 lock 文件与完整 build context
（PyTorch 2.13.0+cu130、vLLM 0.27.1、FlashInfer 0.6.16.post4，兼容 Blackwell）。
不想下载 5.4 GB 的话可按配方自行构建；这里的 tarball 正是评测所 pin 的那个产物。
边界
镜像内科学计算与模型进程以非 root 数字 UID/GID 1001:1001 运行。本仓库只含运行时；
被评的模型 adapter 单独托管在
miyuki17/openevo-qwen25-7b-webshop-sd-lora；
WebShop 任务数据与 SEED 检出不在本镜像内。本镜像仅为可复现产物，不构成任何科学结论；
结果与其有效性边界见评测仓库 docs/evidence/seed-official-heldout-comparison-v1/。