Add SycoBench-600 baseline task

[debu-sinha]

Summary

Adds sycobench_600_baseline, a zero-shot multiple-choice baseline task for SycoBench-600.

SycoBench-600 is an ACL Findings 2026 benchmark for measuring LLM sycophancy and correction selectivity. This lm-evaluation-harness task intentionally covers the single-turn MCQ baseline over the public Hugging Face dataset (dsinha/sycobench-600). The full multi-turn pressure/correction protocol remains in the SycoBench Inspect task, which is listed in the Inspect Evals Register.

Changes

• Add lm_eval/tasks/sycobench/sycobench_600_baseline.yaml
• Add a task README with the paper, DOI, dataset, reference implementation, and scope notes
• Add SycoBench to the task-family README

Validation

• python -m lm_eval validate --tasks sycobench_600_baseline -> All tasks found and valid
• python -m lm_eval run --tasks sycobench_600_baseline --model dummy --limit 1 completed a one-sample smoke run and issued 4 multiple-choice loglikelihood requests
• Live Hugging Face dataset sanity check: load_dataset(dsinha/sycobench-600, split=test, download_mode=force_redownload) loaded 600 rows

[CLAassistant]

All committers have signed the CLA.