BigCode Evaluation Harness
Evaluation harness for code language models with task-specific generation and execution workflows.
External resource. No execution or independent verification is claimed here.
Can a chosen task adapter preserve generation settings, sample identity and execution status through evaluation?
Use synthetic candidate outputs and independently known statuses before evaluating real model generations in a sandbox.
What you could produce
- A version-pinned protocol stating inputs, rights, expected behavior, tolerances and resource limits before execution.
- A retained per-case result and failure ledger with an independently controlled comparison, if qualified execution is later authorized.
Before you use it
- Python, model-generation dependencies and task execution environments; arbitrary generated programs require isolation.
- A separately qualified isolated runtime with a reviewed, pinned dependency and input closure.
Limits to keep in view
- No source program, example, build hook, package, dataset, model or generated research code has been executed or downloaded as a payload.
- The documented self-contained Python 3.13, 90-second pilot does not establish support for this package, its compiled dependencies, GPUs, services or agent sandboxes.
- Installation success, scientific outcomes, runtime compatibility, latency, memory use, API costs and security properties are unmeasured.
Source and permission context
Preserve the upstream project name, repository, exact source commit and applicable contributor/notices; resolve upstream citation guidance for any later formal use.
License evidence recorded. Review the scope and upstream conditions before reuse.
code · Apache-2.0
Apache License version 2.0 is identified in the pinned root license text. Observation is limited to LICENSE at commit 8fc5bae6479c4fbbb28c3f8b644f6a15b3f3b5bd; this is not blanket artifact clearance.
Inspect the license evidence ↗Still unresolved
- Only the cited license and README texts were observed; file exceptions, dependency closure, vendored components and submodules are not comprehensively audited.
- Dataset files, task prompts, generated outputs, model weights, tokenizer assets and hosted APIs are not cleared by a root code license.
- README/documentation reuse rights and version-specific citation guidance remain separately unresolved; only links and original descriptions are retained.
- Harness licensing does not establish rights for linked programming tasks, datasets, repository snippets or model weights.