Add SORRY-Bench dataset loader

[panditdhamdhere]

Summary

Adds a sorrybench dataset source (roadmap Phase 1) so sweeps, campaigns, and leaderboards can pull from the SORRY-Bench 44-class safety taxonomy via source="sorrybench".

Changes

• wallbreaker/datasets/sorrybench.py — loader + normalize to {id, behavior, category, source, benign}
• wallbreaker/datasets/__init__.py — register sorrybench
• wallbreaker/prompts.py — document in agent source= battery list
• tests/test_datasets.py — registry, normalize, and cached battery tests

Notes

• Uses the public SillyTilly/SorryBench CSV mirror (sorry_bench_202406.csv).
• The official sorry-bench/sorry-bench-202503 HuggingFace dataset is gated and requires access approval.
• Categories are prefixed sorry-{1..44} for stratified sampling.

Test plan

• pytest tests/test_datasets.py -q (9 passed)
• Download URL returns 200 (~490 prompts)

[twaai]

thanks for the contribution! merging.

[Gitanuj993]

👍 No harmful and No unwanted code detected !,
It adds a new dataset called sorrybench

[Gitanuj993]

👍 No harmful and No unwanted code detected !,
It adds a new dataset called sorrybench