feat(eval): add 30 multi-turn research use cases dataset

[gspencergoog]

Summary

Adds a new evaluation dataset, eval/datasets/research_use_cases.yaml, containing 30 multi-turn domain research test cases with domain system prompts and qualitative LLM-as-a-judge scoring rubrics.

Changes

• New Evaluation Dataset: Created eval/datasets/research_use_cases.yaml containing 30 translated and de-identified multi-turn research interaction scenarios.
• De-Identification: Anonymized all private personal identifiers, real-world author/student thesis details, family relationships, hardware brand specifics, and local municipal names across all 30 test cases while preserving the underlying domain reasoning challenges.
• Grading Rubrics: Structured qualitative multi-line target grading rubrics for each scenario, specifying expected UI component hierarchy, required data points, and penalization criteria for LLM-as-a-judge scoring.
• Encryption: Staged and encrypted the dataset file using Transcrypt.

Impact & Risks

• No breaking changes: This is an additive dataset in eval/datasets/.
• Presubmit evaluation load: The dataset will be automatically included in standard evaluation runs that scan eval/datasets/*.yaml.

Testing

• Validated dataset schema conformance with jsonschema.validate against eval/datasets/dataset_schema.json.
• Executed uv run pytest tests/test_dataset.py in eval/ (5/5 passed).
• Executed a single-epoch test evaluation via uv run python main.py --dataset research_use_cases --epochs 1 confirming successful execution and logging.
• Verified Transcrypt ciphertext encryption on Git object via git cat-file -p :eval/datasets/research_use_cases.yaml.

[gemini-code-assist]

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.