feat(tests): add EvalScope evaluation suite with needle-in-haystack and acc test

[Potterluo]

Summary

This PR introduces a comprehensive evaluation suite built on EvalScope (v1.5.2) to automate accuracy testing for LLMs. It supports two primary evaluation modes:

• Mainstream Benchmarks: aime24, aime25, aime26, gsm8k, longbench_v2, ceval, cmmlu, humaneval, mmlu, mmlu_pro, etc.
• Needle In A Haystack: Long-context retrieval evaluation with configurable context length ranges.

Key Changes

1. New Utility Class: EvalScopeRunner

• Encapsulates run() and collect_results() logic.
• Handles timestamp-based result directory discovery and JSON report parsing.
• Exports metrics via the existing @export_vars decorator for database storage.

2. Refactored Test Cases (test_evalscope.py)

• Configuration building moved to local helper functions (_build_general_task_config, _build_needle_task_config).
• Environment variable support:
  • SCOPE_DATASET_ROOT / SCOPE_TREST_LIST
  • SCOPE_NEEDLE_MIN / SCOPE_NEEDLE_MAX
• Clear separation between test logic and infrastructure code.

3. Enhanced Needle Content

• Replaced the public "San Francisco sandwich" needle with a fictional, unique passage to avoid training data contamination.
• Added corresponding Chinese needle for bilingual subset testing.

4. Documentation (README.md)

• Bilingual (Chinese/English) guide covering:
  • Environment setup and dataset preparation (online/offline).
  • Configuration variables and local override methods.
  • Execution commands and result output structure.
  • Example JSON output and screenshot references.

How to Test

cd test

# Mainstream benchmarks
pytest suites/E2E/test_evalscope.py::test_eval_accuracy

# Needle In A Haystack
pytest suites/E2E/test_evalscope.py::test_needle_task

# All EvalScope tests
pytest --feature=evalscope

[yuanzhg078]

typo here : SCOPE_TREST_LIST