Datasets:

lolbench26
/

LoLBench

Prefer the published Docker Hub image when it is available. Harbor refers to the image as smartdub26/lolbench-linux:cpython_2-1.0.0, so Docker can pull it directly when it is not present locally:

docker pull smartdub26/lolbench-linux:cpython_2-1.0.0

Only if the Docker Hub image is unavailable, build the single task image locally:

docker build -t lolbench/cpython_2:1.0.0 harbor_tasks/cpython_2/environment
docker tag lolbench/cpython_2:1.0.0 smartdub26/lolbench-linux:cpython_2-1.0.0

The image contains the entire processed CPython source tree at the base commit under /workspace/cpython; Harbor must not mount a host source checkout for the agent.

The agent and verifier both use this same image. Private verifier files are not baked into the image; Harbor mounts them only in the separate verifier phase.

Run an Agent

Use Harbor to run an agent. Replace the agent and model with the evaluation configuration you want.

.harbor-venv/bin/harbor run \
  -p harbor_tasks/cpython_2 \
  -a codex \
  -m "<model>" \
  --job-name cpython_2_agent_union \
  --jobs-dir harbor_runs/cpython_2 \
  --no-delete \
  --n-concurrent 1 \
  --ve LOLBENCH_SUITE=union

The agent works in:

/workspace/cpython

After implementing the requirement, the agent must run:

lolbench-submit

This writes:

/logs/artifacts/solution.patch

On the host, that patch is available at:

find harbor_runs/cpython_2/cpython_2_agent_union -path '*/artifacts/solution.patch' -print

Evaluate a Patch

The verifier consumes /logs/artifacts/solution.patch and writes a sanitized report plus reward. Choose one mode with LOLBENCH_SUITE=orig, aug, or union; if unset, the task defaults to union. To evaluate an existing patch through Harbor, run the oracle agent as a patch loader and mount the candidate patch into the agent container:

PATCH="$(pwd)/path/to/solution.patch"

.harbor-venv/bin/harbor run \
  -p harbor_tasks/cpython_2 \
  -a oracle \
  --job-name cpython_2_eval_union \
  --jobs-dir harbor_runs/cpython_2 \
  --no-delete \
  --n-concurrent 1 \
  --mounts '[{"type":"bind","source":"'"$PATCH"'","target":"/candidate/solution.patch","read_only":true}]' \
  --ae LOLBENCH_ORACLE_PATCH=/candidate/solution.patch \
  --ve LOLBENCH_SUITE=union

Outputs:

find harbor_runs/cpython_2/cpython_2_eval_union -path '*/verifier/agent_report.json' -print
find harbor_runs/cpython_2/cpython_2_eval_union -path '*/verifier/reward.json' -print

Use LOLBENCH_SUITE=orig or LOLBENCH_SUITE=aug with separate output job names to evaluate the other modes.

Notes

The source in environment/repo/ is an archived base-commit snapshot. It does not contain CPython's original .git directory. During build, the Dockerfile first removes any /workspace/cpython checkout inherited from the base image, then copies this snapshot and creates a fresh local git repository with a single synthetic base commit. Agents can produce a normal patch without seeing or recovering later upstream commits through git.

tests/docker-compose.yaml is a verifier-only Harbor overlay. It mounts tests/, tests/private/, and solution/ into the separate verifier container; none of those paths are mounted into the agent container.

solution/solution.patch is the maintainer oracle patch for validation. Do not mount it into agent containers.