Creates Judge Example as a wrapper on Policy

[Jack-Khuu]

Update(10/10) : This is a very much simplified version of the previous iterations of the PRs. It provides just a basic example of LLM Judges for GRPO

Judges can both be used as "Verifiers" or "Graders". This PR adds to the sandbox, a CorrectnessJudge example of how an LLM Judge can be used in GRPO (note that this PR does not integrate)

It takes as input (prompt + response) generated from a model, and returns whether the model thinks it accurately responded to the prompt. Results can then be used to make decisions during GRPO whitening

---

python -m tests.sandbox.vllm.judge --config tests/sandbox/vllm/qwen3_4b.yaml

[allenwang28]

Prompt: What is the capital of Japan?
Responses: ['Aardvark', 'Durian', 'Tokyo']

Generation Results:
================================================================================
Sample 1
Evaluation: 3
--------------------------------------------------------------------------------
Sample 2
Evaluation: 3
--------------------------------------------------------------------------------
Sample 3
Evaluation: 3
--------------------------------------------------------------------------------
Sample 4
Evaluation: 3
--------------------------------------------------------------------------------

lol is this working correctly?

[Jack-Khuu]

lol is this working correctly?

I wrote this prompt from the deep archives of my mind and I'm also shocked that the prompting worked.

[Jack-Khuu]

Update (10/10): This is a very much simplified version of the previous iterations of the PRs. It provides just a basic example of LLM Judges for GRPO

There is out of scope future work leveraging structured decoding, but that requires additional investigation on how to configure with CoT models (which need to think)