LLMs are probabilistic systems. Reproducibility of completions is not guaranteed. It only makes sense to account for this, by running challenges multiple times to obtain a success ratio rather than a boolean success/failure result.
Changes:
- Add `-N`, `--attempts` option to CLI and `attempts_per_challenge` parameter to `main.py:run_benchmark`.
- Add dynamic `i_attempt` fixture through `pytest_generate_tests` hook in conftest.py to achieve multiple runs per challenge.
- Modify `pytest_runtest_makereport` hook in conftest.py to handle multiple reporting calls per challenge.
- Refactor report_types.py, reports.py, process_report.ty to allow multiple results per challenge.
- Calculate `success_percentage` from results of the current run, rather than all known results ever.
- Add docstrings to a number of models in report_types.py.
- Allow `None` as a success value, e.g. for runs that did not render any results before being cut off.
- Make SingletonReportManager thread-safe.
* feat(benchmark): Add JungleGym WebArena challenges
- Add `WebArenaChallenge`, `WebArenaChallengeSpec`, and other logic to make these challenges work
- Add WebArena challenges to Pytest collection endpoint generate_test.py
* feat(benchmark/webarena): Add hand-picked selection of WebArena challenges
