FEAT: Add Anthropic model-written evals dataset

[Haquant (Steeve-Crypto)]

Description

Adds the Anthropic model-written evals dataset as a remote seed dataset loader, so it can be fetched through SeedDatasetProvider.

The loader reads the full release from https://github.com/anthropics/evals. The Hugging Face dataset only publishes a subset, which is why this does not use load_dataset.

Supported collections are persona, sycophancy, advanced-ai-risk (human-written and LM-written eval files), and winogenerated examples. Few-shot generator prompts and the occupation catalog are left out because they are not eval items.

Rows with a question use that text as the seed prompt. answer_matching_behavior and answer_not_matching_behavior are kept in seed metadata, including when the non-matching field is a list. Winogenerated examples have no question column, so the prompt is sentence_with_blank and pronoun options stay in metadata.

This follows the current _RemoteDatasetLoader pattern. Closed PR #1170 targeted an older question-answering layout and was not merged.

Content warning: these prompts are meant to provoke model behavior and may contain offensive content.

Fixes #450

Tests and Documentation

Unit tests cover category filtering, behavior-label metadata, list-valued answers, winogenerated fill-in sentences, skipped non-eval files, GitHub listing errors, and an empty result. They use mocked responses and do not download the dataset.

Commands:

• uv run pytest tests/unit/datasets/test_anthropic_model_written_evals_dataset.py (8 passed)
• uv run pytest tests/unit/datasets/test_seed_dataset_provider.py -k _AnthropicModelWritten (metadata registration passed)
• uv run ruff check and uv run ruff format --check on the changed files
• uv run ty check pyrit/datasets
• SKIP=ty-check uv run pre-commit run --files on the changed files

The pre-commit ty hook typechecks all of pyrit with --extra all and was not run. ty check on pyrit/datasets passed.

Jupytext was not run. The dataset is discovered through SeedDatasetProvider, so no notebook sample was added.

No CLA was signed from this change.

[Haquant (Steeve-Crypto)]

@microsoft-github-policy-service agree

[Haquant (Steeve-Crypto)]

Roman Lutz (@romanlutz) Addressed all four review comments in 05d9fff. Ready for another look.

[Haquant (Steeve-Crypto)]

Roman Lutz (@romanlutz) Thanks for the approval. The merge queue failed twice on test_sqlite_deadline_bounds_file_database_lock_wait_and_restores_timeout (Windows only, took ~1.2s vs the 1s limit). This PR doesn't touch memory code, so it looks flaky. Mind re-queueing when you get a chance?

[Roman Lutz (romanlutz)]

I'm fixing that in #3121 hopefully but need approval myself...