Add LEXam legal exam benchmark to Swiss legal evals (#1243)
Adds LEXam (https://huggingface.co/datasets/LEXam-Benchmark/LEXam) tasks:
- lexam_oq:{en,de}: open-ended legal exam questions scored with an
LLM-as-judge (JudgeLEXamOQ, default openai/gpt-4o-2024-11-20 via LiteLLM).
- lexam_mcq_{4,8,16,32}{,_idk}:{en,de}: letter-based MCQ scorer using
IndicesExtractionConfig with a ###X### / \boxed{X} / "Final answer: X"
fallback regex. Choices are shuffled deterministically (seed 42).
- Optional "I don't know" calibration: when with_idk=True, an extra letter
slot is reserved for IDK and the metric reports trad_score, idk_score
(+1/0/-1), idk_freq, and extract_fail; otherwise it reports acc and
extract_fail.
Adds 18 tasks to TASKS_TABLE (2 OQ x 2 langs and 4 choice-counts x 2 langs
x {plain, IDK}).
Co-authored-by: Joel Niklaus <joel@hf.co>
Co-authored-by: Cursor <cursoragent@cursor.com>
