zhuiluozhiji/DRBenchPro-runtime

DRBenchPro Runtime

Release status: Ready

This repository distributes the standalone linux/amd64 Docker image for DRBenchPro, including the application, benchmark, and fixed English Wikipedia SQLite/FTS5 search index. Model weights and API credentials are supplied separately by users.

The benchmark's browsable queries and cases are in the separate DRBenchPro dataset. Users evaluating their own generated reports can use the Python evaluator without downloading this large Wikipedia runtime; see the code quickstart.

Release

Item	Value
Image tag after import	drbenchpro/standalone:0.3.0-enwiki-20260701-v1-23db51b
Image ID	sha256:1b49b90c9646d78e6a676c148534f3c9407a5f8c8d6a0a98cada1283e0cc830a
Platform	linux/amd64
Source revision	23db51b
Wikipedia snapshot	enwiki-20260701-v1
Image size before archive compression	172,187,227,486 bytes (about 160.36 GiB)
Compressed download	63,018,778,599 bytes (58.69 GiB)
Parts	12; each up to 5 GiB (5,368,709,120 bytes)
Format	One gzip-compressed docker save archive, split into numbered parts
Archive payload revision	ad17787eee5750fdc39775d6d4f6c26e5c097563

All parts of this release are required. Their order, sizes, SHA-256 values and complete archive SHA-256 are recorded in release.json. SHA256SUMS checks individual parts; ARCHIVE_SHA256 checks the ordered combined gzip stream. These hashes serve different purposes from the embedded database hash.

Requirements and disk space

• A Linux amd64 Docker environment, plus the Docker CLI. Native ARM operation has not been validated; Apple Silicon requires amd64 emulation.
• The Hugging Face hf CLI for batch downloading. The provided helpers require Python 3.10+ and Bash.
• Space for the 58.69 GiB download, in addition to Docker's image storage and temporary extraction needs. About 350 GiB of free space in the Docker filesystem, plus the compressed download if retained on that same filesystem, is a planning allowance rather than a measured minimum. Check physical host capacity and, for Docker Desktop, the Linux VM disk capacity.
• No GPU is required when using external model APIs. Minimum CPU/RAM and real-provider throughput have not been established.

The bundled database is 172,016,893,952 bytes; this is not a RAM requirement. The import helper streams the parts directly to Docker, so it does not create an additional combined archive file.

Download all parts

Run this after the release status above is Ready:

hf download zhuiluozhiji/DRBenchPro-runtime \
  --repo-type dataset \
  --revision ad17787eee5750fdc39775d6d4f6c26e5c097563 \
  --include 'images/0.3.0-enwiki-20260701-v1-23db51b-linux-amd64/*' \
  --include 'scripts/*' \
  --local-dir DRBenchPro-runtime

The fixed revision keeps all parts and helpers from one release together. The current dataset card on this page reports the final release status. If interrupted, rerun the same command with the same revision and local directory. Preserve the download metadata/cache until the download completes. This public, ungated repository can be downloaded without an account token.

If the helpers are already available, they can also download the release:

bash DRBenchPro-runtime/scripts/download.sh \
  --revision ad17787eee5750fdc39775d6d4f6c26e5c097563 \
  --directory DRBenchPro-runtime

Verify and import into Docker

bash DRBenchPro-runtime/scripts/verify.sh
bash DRBenchPro-runtime/scripts/load.sh

verify.sh reads all parts and checks sizes, order, individual hashes and the complete archive hash. load.sh repeats verification before sending the ordered gzip stream to docker image load, then checks the loaded image ID and platform and restores the tag shown above. It stops on corrupt/missing parts or a conflicting local image tag. Verification of this large archive takes time.

docker save by image ID can omit the original tag in its archive metadata; restoring the documented tag after verifying the image ID is intentional.

export DRBP_IMAGE=drbenchpro/standalone:0.3.0-enwiki-20260701-v1-23db51b
docker image inspect "$DRBP_IMAGE" --format '{{.Id}} {{.Os}}/{{.Architecture}}'

Expected identity:

sha256:1b49b90c9646d78e6a676c148534f3c9407a5f8c8d6a0a98cada1283e0cc830a linux/amd64

Start the bundled Wikipedia service

docker run -d --name drbp-standalone \
  --read-only --tmpfs /tmp:rw,nosuid,nodev,size=256m \
  -p 127.0.0.1:18091:8091 \
  "$DRBP_IMAGE"

docker inspect --format '{{.State.Health.Status}}' drbp-standalone
docker logs drbp-standalone

Wait for healthy. Default startup serves the bundled Wikipedia index and does not call Research or Judge models. Do not mount another directory over /opt/drbenchpro/wiki. Containers share the image's read-only data layers.

The container includes full usage instructions and blank API configuration examples. Read them without starting model calls:

docker run --rm --network none --entrypoint cat \
  "$DRBP_IMAGE" /app/docker/README.md

docker run --rm --network none "$DRBP_IMAGE" verify-data
docker run --rm --network none "$DRBP_IMAGE" validate

For Research, external-report evaluation, Judge, metrics and resume commands, see the frozen container guide. Use this release's DRBP_IMAGE tag in those commands. Research requires MODEL_API_KEY, MODEL_BASE_URL, and MODEL_NAME; Judge requires JUDGE_API_KEY, JUDGE_BASE_URL, and JUDGE_MODEL. Supply your own values at runtime. Research/Judge execution can incur charges from your model provider. API credentials are not included in this image.

Validation scope

The frozen image passed 142 native checks, Wikipedia Search/Open/Find smoke checks, a full Research CLI workflow using real Wikipedia with mock models, and independent evaluation of 288 synthetic external reports with 88,204 rubric scores and matching reference metrics. These were software execution checks with zero paid model calls; they do not establish model quality.

Distribution validation: Passed: full anonymous download using an independent cache, every part and combined archive SHA-256, gzip integrity, Docker import with exact image ID/platform, benchmark validation, and Wikipedia Search/Open/Find. Import was tested on a host that already held the image and could reuse existing Docker layers; a fresh-host first install has not been validated.

Sources and licenses

The image includes English Wikipedia-derived content, manifests, attribution notices and license texts. See NOTICE.md and the files embedded under /opt/drbenchpro/licenses/ and /opt/drbenchpro/wiki/WIKIPEDIA_CONTENT_NOTICE.md. Retain applicable Wikipedia attribution and license notices when redistributing derived content.

Code and benchmark license decisions are pending, as stated in the source repository. Wikipedia and third-party licenses apply to their respective components; they do not automatically license the DRBenchPro code or benchmark.

For paper and project updates, use the DRBenchPro code repository.