feat: add AIME-2026 benchmark.

[xxman-google]

What does this PR do ?

Add AIME'26 eval benchmark.

Many new models just came out this year may incorporated data from 2025. I tested the new benchmark with Qwen3.5-9B model and it scored 89.2% averaging 16 runs.

Issues

N/A

Usage

• Launch eval through the following command:

uv run examples/run_eval.py \
data.dataset_name="aime2026" \
data.prompt_file=examples/prompts/cot.txt \
cluster.gpus_per_node=8 \
generation.vllm_cfg.max_model_len=81920 \
generation.model_name="Qwen/Qwen3.5-9B" \
generation.temperature=1.0 \
generation.top_p=0.95 \
generation.top_k=20 \
eval.num_tests_per_prompt=16

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yuki-97]

thanks for adding this. overall lgtm! one minor comment.

could you also help to add an unit test for the dataset like those in tests/unit/data/datasets/test_eval_dataset.py?

[xxman-google]

Comments addressed. PTAL @yuki-97

[yuki-97]

lgtm, and offline verified test_aime_dataset works (since it's skipped now).

[yuki-97]

/ok to test ba1b343

[yuki-97]

hi @xxman-google , could you help to fix the lint fail? other fails are not related to this PR, we'll take a look.

[xxman-google]

hi @xxman-google , could you help to fix the lint fail? other fails are not related to this PR, we'll take a look.

@yuki-97 Done. PTAL again.

[yuki-97]

/ok to test 89cdaa5

[yuki-97]

/ok to test 5c32bac

[yuki-97]

@kajalj22 could you help to take a look? I've tried both Lfast and L1, but both of them fail when building container (different reason).