evaluation_code/external_benchmarks/README.md

External benchmark adapters

The repository does not redistribute third-party benchmark images, metadata, or evaluation implementations. Pin an upstream commit for every reported result and record that commit in the release notes.

MME-CoF

scripts/eval_mme_cof.sh implements the five-aspect VLM evaluation used by the current experiment code. It requires a JSON array whose records contain at least id, user_prompt, and task_name.

Before release, reconcile the local experiment metadata with the paper:

• paper description: 59 samples;
• current local samples_origin.json: 55 samples.

Do not label either version canonical until this difference is resolved.

RULER-Bench

scripts/eval_ruler_bench.sh prepares one directory per seed and task. It does not run or vendor the upstream evaluator. After preparation, invoke the pinned upstream evaluator according to that repository's documented input contract.

Using separate seed directories prevents the previous adapter from silently overwriting multiple seeds with the same target filename.

V-ReasonBench

scripts/eval_v_reasonbench.sh converts generated video names to the format consumed by the upstream evaluator. It does not vendor the upstream code.

Before release, reconcile the local experiment metadata with the paper:

• paper description: 327 samples across 12 categories;
• current local metadata: 328 samples across 13 categories.

The old local evaluator also ran only block_slide; it must not be presented as a complete V-ReasonBench evaluation.