Samples / README.md
TidorP's picture
Review selected tasks, correct cyber reference cache identity, and clarify reproducibility setup
ddf04fa verified
|
Raw History Blame Contribute Delete
10.5 kB
metadata
license: cc-by-4.0
language:
  - en
  - de
  - ja
  - fr
tags:
  - benchmark
  - evaluation
  - docker
  - stem
  - software-engineering
pretty_name: Poindexter Labs Samples
configs:
  - config_name: gdpval
    data_files:
      - split: sample
        path: data/gdpval/tasks.jsonl
  - config_name: stem
    data_files:
      - split: sample
        path: data/stem/tasks.jsonl
  - config_name: cyber
    data_files:
      - split: sample
        path: data/cyber/tasks.jsonl
  - config_name: qna
    data_files:
      - split: sample
        path: data/qna/tasks.jsonl

Poindexter Labs Samples

Seven selected tasks in one Docker package: three localized professional-work tasks, two STEM problems, and two repository tasks. Researchers bring their own API keys. Native OpenAI and Anthropic are supported; no Azure or AWS account is required.

Sample Language / type Source
german-664a42e5-3240-413a-9a57-ea93c6303269 German · family-foundation presentation GDPVal Select 25+25
japanese-ec2fccc9-b7f6-4c73-bf51-896fdb433cec Japanese · photography NFT article GDPVal Select 25+25
french-43dc9778-450b-4b46-b77e-b6d82b202035 French · income-tax return GDPVal localization project
PDS136 STEM · algebra and topology Block Picard Groups and Surface Bundles
PDS049 STEM · organic chemistry Microwave Multicomponent Product Formula
swe_rl_cybersecurity_task_02 Code · certificate-chain validation cyber_02
pdl_qna_signed_package_resolution_014 Code Q&A · signed package resolution qna014

For a readable introduction, see STEM Data Samples.pdf. Prompts, rubrics, and worked solutions are also available as PDS136 and PDS049. The GDPVal reading copy contains the localized prompts, input links and rubrics. The Q&A reference explains the expected audit. Task records are in data/*/tasks.jsonl. Input documents and repository archives are included; no Poindexter service account or production connection is needed.

Quick start

Install Docker and Python. Clone this dataset with Git LFS, or install the Hugging Face CLI and download it:

python -m pip install --upgrade huggingface_hub
hf download Poindexter-Labs/Samples --repo-type dataset --local-dir samples
cd samples
docker build -t poindexter-samples .
cp .env.example .env
# Edit .env to add your own keys and available model IDs.
mkdir -p results
docker run --rm poindexter-samples list
docker run --rm poindexter-samples plan --group all

Run the two STEM samples through native OpenAI:

docker run --rm --env-file .env -v "$PWD/results:/results" \
  poindexter-samples run --group stem --provider openai

Use native Anthropic instead:

docker run --rm --env-file .env -v "$PWD/results:/results" \
  poindexter-samples run --group stem --provider anthropic

--model MODEL_ID overrides the solver model. --judge-model MODEL_ID overrides the native OpenAI rubric judge. Anthropic solver runs still need OPENAI_API_KEY for grading. The GDPVal runner supports gpt-6-astra, gpt-6.2-astra, claude-opus-5-5, and claude-fable-5-2 with their recorded context/output budgets. Other groups accept custom model IDs. The default solver IDs are examples from our selected model configurations; choose IDs available to your own account. One attempt per task runs by default; use --attempts 8 for eight attempts. Requests and sandbox use are billed to your accounts.

Run all seven

Before running all seven, you need:

Setting Used for
OPENAI_API_KEY and ANTHROPIC_API_KEY Native model calls and the two-provider Q&A judge panel
Model IDs available to your accounts Solver and judge settings in .env.example
E2B_API_KEY and GDPVAL_E2B_TEMPLATE_ID GDPVal document-execution environment
BRAVE_API_KEY GDPVal web research
Running Docker and the socket mount below Code-task sandboxes and evaluators

GDPVal uses the same E2B environment as the public Mini-Bench package. It additionally needs E2B_API_KEY, GDPVAL_E2B_TEMPLATE_ID, and BRAVE_API_KEY. If you do not already have an appropriate template, build it once in your own E2B account:

docker run --rm --env-file .env poindexter-samples template
# Put the printed template ID into GDPVAL_E2B_TEMPLATE_ID in .env.

The template includes office, PDF, font, and document-processing tools and can take some time to build. Its build script and dependency inventory are included under engines/general/vendor/tools/gdpval/.

The code tasks use local Docker sandboxes and evaluators, controlled through the Docker socket. This socket grants the controller access to your Docker daemon; use a trusted local machine. The model's sandbox does not receive the socket, API keys, hidden tests, reference patches, or grading rubrics. The controller stages only solver-visible inputs. Keep /results bind-mounted so child containers can access the submitted work.

docker run --rm --env-file .env \
  -v "$PWD/results:/results" -v /var/run/docker.sock:/var/run/docker.sock \
  poindexter-samples doctor --group all

docker run --rm --env-file .env \
  -v "$PWD/results:/results" -v /var/run/docker.sock:/var/run/docker.sock \
  poindexter-samples run --group all --provider openai

This is one controller image and entry point. Code execution starts isolated child containers; GDPVal execution runs in E2B. It is not an offline, all-in-one sandbox. Each group runs sequentially. doctor verifies local inputs, imports, configuration and Docker setup; it does not contact model APIs. plan validates task files and prints the required settings without starting episodes or calling providers.

Use --group gdpval|stem|cyber|qna to run a subset, or --task EXACT_ID (repeatable) to select individual samples. Unknown IDs and conflicting group selections are rejected before execution.

Results and retries

Results remain in results/{group}/: attempt records, provider/model identities, task hashes, answers or patches, grading feedback and reports. Run the same command again to resume; completed attempts are preserved. A grading failure can be retried without generating a new answer:

docker run --rm --env-file .env \
  -v "$PWD/results:/results" -v /var/run/docker.sock:/var/run/docker.sock \
  poindexter-samples grade --group all --provider openai

docker run --rm -v "$PWD/results:/results" poindexter-samples report --group all

Use the same provider/model/options when resuming or grading. A changed configuration is recorded separately. --retry-failed explicitly permits a failed solver attempt to run again. Technical failures stay ungraded and are not counted as incorrect model answers.

GDPVal submissions are limited to 20 files, 10 MiB per file and 25 MiB total by the inherited runner. Oversized output is rejected before grading. The French live smoke run produced a 19.3 MB PDF and hit this limit; its smaller supplied human reference graded successfully. See the validation record for this remaining execution limitation.

Scoring and provenance

  • GDPVal: the selected localized prompts, rubrics and inputs are preserved. The rubric evaluator reports a normalized score and the Mini-Bench 95% pass threshold. The French reference deliverable is included under data/gdpval/references/ for human inspection; it is not supplied to the solver.
  • STEM: the updated November 18, 2026 PDF supplies the prompts and weighted rubrics: 10 points for PDS136 and 7 for PDS049. The chemistry rubric's two partial-credit criteria are represented as four one-point checks, preserving its 1/1/1/2/2 grouping. Reports give the normalized weighted score; passed means full rubric credit (100%) for these samples. This is not the old answer-only STEM benchmark protocol. The PDF's historical chart is context, not a score reproduced by these revised tasks.
  • cyber_02: the reference agent loop submits a patch, graded by the task's public and hidden tests under the published scoring_v0.2 protocol; the pass rule is strictly greater than 0.95.
  • qna014: the reference agent loop submits answers/answer.md, graded with the published scoring_llm_v0.3 protocol. For easy native-key use, Samples defaults to two native judges (OpenAI and Anthropic), not the original three-provider panel. QnA requires both API keys and at least two judge members. Set SAMPLE_ANTHROPIC_JUDGE_MODEL to change the Anthropic judge. The actual panel is recorded. Supply --judges 'openai:MODEL@https://api.openai.com/v1,anthropic:MODEL' to use another panel. Such scores are not directly comparable with the historical panel's scores.

The repository tasks retain the published 2M-token profiles, including their long episode budgets. A single task can therefore take substantially longer than one model API call. See each plan's budgets before starting.

Runner code and repository tasks originate from Poindexter-Labs/PDL-Bench, pinned to d68375b39ff4e7dc69e9ee9107d475d3ee9808b4. SOURCE_MANIFEST.json records the selected inputs and packaging adaptations; CHECKSUMS.json covers this release. Vendored notices and licenses are retained under engines/. Separate Python environments preserve the two upstream runners' dependency pins inside one image.

For validation results and their limits, see VALIDATION.md. The five-task accuracy review records the independent checks and the corrected cyber reference. Documentation, input files, references and offline checks are also included inside the controller at /opt/samples.

Reference baselines

To exercise the executable task without model calls:

docker run --rm -v "$PWD/results:/results" \
  -v /var/run/docker.sock:/var/run/docker.sock \
  poindexter-samples run --group cyber --agent gold

--agent noop tests an unchanged repository. Both baselines use the same submission and evaluation path. A Q&A gold baseline still calls the configured judge. Gold patches and worked solutions are included for reproducibility and human review, not model input.

License

Task data is distributed under CC BY 4.0. Original GDPVal attribution: OpenAI, with localized adaptations by Poindexter Labs. Vendored software retains the licenses and notices distributed with the source package.