[gwarmstrong]

Add mbpp benchmark

Summary

Migrates the MBPP+ benchmark (378 Python function-completion tasks)
from NeMo Skills into Gym as a data-only follow-up to the HumanEval+
migration (#1201). Reuses the evalplus resource server unchanged —
its dataset: humaneval | mbpp knob was already in place — so this
PR adds only the new benchmarks/mbpp/ directory.

Includes

• benchmarks/mbpp/ — benchmark dataset binding
  • Data source: MBPP+ via evalplus.data.get_mbpp_plus()
    (378 tasks, MIT license).
  • prepare.py mirrors NeMo-Skills' nemo_skills/dataset/mbpp/prepare.py:
    same evalplus source, same per-row prompt.replace(" ", "\t")
    (Skills comment: "models like tabs more than spaces"), drops
    base_input / plus_input (re-fetched at eval time by EvalPlus).
  • prompts/default.yaml is character-for-character identical to
    Skills' nemo_skills/prompt/config/generic/codegen.yaml.
  • config.yaml inherits the evalplus resource server with
    dataset: mbpp; reuses the evalplus_simple_agent configuration.

Validation (data-only / Mode 2)

Skills-vs-Gym prepare-output parity, run on the same evalplus==0.3.1
upstream:

Skills file: nemo_skills/dataset/mbpp/test.jsonl
Gym file:    benchmarks/mbpp/data/mbpp_benchmark.jsonl

Row count:   378 (Skills)  ==  378 (Gym)
Field-level mismatches across 378 rows on the 4 retained fields
(question, entry_point, canonical_solution, task_id):  0

Documented schema rename: Skills' top-level task_id is moved into
verifier_metadata.task_id to match the evalplus resource server's
verifier contract; Skills' extra informational fields (assertion,
atol, contract, prompt) are dropped (not consumed by either
evaluator). This mirrors the schema convention adopted in the
benchmarks/human_eval/ sibling.

The evalplus resource server's existing test suite (24 tests,
including TestLoadDatasetAndExpected.test_mbpp_path) continues to
pass without modification.

Licensing

MBPP+ is MIT-licensed (https://github.com/evalplus/evalplus).

[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.