feat(tasks): add public model benchmark tags

[Luodian]

Summary

This PR adds curated tag metadata that maps runnable lmms_eval image tasks to model families whose official public evaluation materials highlight those benchmarks.

The goal is to make it possible to select a practical, vendor-aligned benchmark slice directly from YAML using existing TaskManager tag expansion, without introducing any new evaluator logic.

What changed

This PR adds four model-family tags:

• public_eval_qwen3_5_family
• public_eval_seed2_family
• public_eval_gemini3_family
• public_eval_gpt5_family

It applies those tags to the corresponding task YAMLs and adds documentation for the naming convention, scope, and maintenance rules in docs/guides/model_benchmark_tags.md.

Current tag expansion

public_eval_qwen3_5_family

• ai2d -> lmms-lab/ai2d (test)
• charxiv_val_reasoning -> princeton-nlp/CharXiv (validation)
• hallusion_bench_image -> lmms-lab/HallusionBench (image)
• mathvista_testmini_cot -> AI4Math/MathVista (testmini)
• mathvista_testmini_format -> AI4Math/MathVista (testmini)
• mathvista_testmini_solution -> AI4Math/MathVista (testmini)
• mmbench_en_dev -> lmms-lab/MMBench config en (dev)
• mmlongbench_doc -> yubo2333/MMLongBench-Doc (train)
• mmmu_pro_standard -> MMMU/MMMU_Pro config standard (10 options) (test)
• mmmu_pro_vision -> MMMU/MMMU_Pro config vision (test)
• mmmu_val -> lmms-lab/MMMU (validation)
• mmstar -> Lin-Chen/MMStar (val)
• ocrbench -> echo840/OCRBench (test)
• omnidocbench -> ouyanglinke/OmniDocBench_tsv (train)
• realworldqa -> lmms-lab/RealWorldQA (test)

public_eval_seed2_family

• charxiv_val_descriptive -> princeton-nlp/CharXiv (validation)
• charxiv_val_reasoning -> princeton-nlp/CharXiv (validation)
• mathvision_testmini -> MathLLMs/MathVision (testmini)
• mmlongbench_doc -> yubo2333/MMLongBench-Doc (train)
• ocrbench_v2 -> ling99/OCRBench_v2 (test)
• omnidocbench -> ouyanglinke/OmniDocBench_tsv (train)

public_eval_gemini3_family

• charxiv_val_reasoning -> princeton-nlp/CharXiv (validation)
• mmmu_pro_standard -> MMMU/MMMU_Pro config standard (10 options) (test)
• mmmu_pro_vision -> MMMU/MMMU_Pro config vision (test)
• omnidocbench -> ouyanglinke/OmniDocBench_tsv (train)

public_eval_gpt5_family

• mmmu_val -> lmms-lab/MMMU (validation)

Why this helps

Today, the repo already supports tag as a first-class task selector, but there is no built-in way to ask for "the public benchmark slice most associated with model family X".

With these tags in place, users can run commands like:

uv run python -m lmms_eval --tasks public_eval_qwen3_5_family

This keeps the implementation simple, avoids adding hard-coded model-specific logic in Python, and makes future curation incremental at the YAML layer.

Validation

• pre-commit run --all-files
• Verified that TaskManager indexes and expands all four new tags successfully