Source packages
This reference describes the TCSExam source package and its build entry points. It is for users preparing environments from source or creating a source distribution.
The repository contains 100 tasks in 130 task folders: 70 shared Lean folders and 30 Solver tasks with separate Online and Offline folders. Lite selects 20 tasks from across Full's task categories, using these same folders.
Read the table by directory. Runtime outputs are kept separately from source.
| Path | Contents |
|---|---|
tasks/ | Instructions, task configuration, environment recipes and verifiers |
suites/ | Full/Lite selections and task/protocol paths |
src/tcsexam/ | CLI, score calculation, judging backend and fixed scoring references |
assets/ | Resource inventories, pinned input recipes, environment builds and notices |
docs/ | Usage, technical references, sources and licenses |
Each task uses Harbor's folder format with instruction.md, task.toml, environment/ and tests/. Runtime images, bulk prepared datasets, model outputs and reference-answer archives are separate products.
Build from source
Follow getting started to install the CLI and check the host. The build entry points are:
tcsexam source-inputs
tcsexam build --suite litesource-inputs prepares fixed public Lean source archives. build prepares the selected dependencies and images on the Docker host. Use resource preparation for fixed inputs, acquisition failures, image manifests and selecting a smaller build. Ordinary evaluation uses the task inputs and verifiers; reference-answer archives are optional inputs for the reference-answer tooling.
To prepare all task variants, build Full once with Online Solver and then the Offline Solver track:
tcsexam build --suite full --solver-protocol online
tcsexam build --suite full --track solver --solver-protocol offlineConfigure the model and check the chosen runtime selection through running evaluations. Online Solver additionally needs its backend service.
Export and redistribution
CONTRIBUTING.md contains source-export commands and development checks. Export keeps source separate from caches, credentials and evaluation outputs. Keep runtime images and job directories separately when preserving an experiment.
Preserve the applicable licenses and notices. Distribution conditions describe which third-party materials and images can be shared.