license:

• apache-2.0
• cc-by-4.0
  task_categories:
• text-generation
  language:
• en
  tags:
• reinforcement-learning
• rlvr
• evaluation
• benchmark
  configs:
• config_name: AIME2025
  data_files:
  • split: test
    path: AIME2025/test-*.parquet
• config_name: AIME2026
  data_files:
  • split: test
    path: AIME2026/test-*.parquet
• config_name: GPQA
  data_files:
  • split: test
    path: GPQA/test-*.parquet
• config_name: LCB_v5
  data_files:
  • split: test
    path: LCB_v5/test-*.parquet
• config_name: LCB_v6
  data_files:
  • split: test
    path: LCB_v6/test-*.parquet
• config_name: IFEval
  data_files:
  • split: test
    path: IFEval/test-*.parquet
• config_name: IFBench
  data_files:
  • split: test
    path: IFBench/test-*.parquet
• config_name: BFCL_v3
  data_files:
  • split: test
    path: BFCL_v3/test-*.parquet

Multi-Domain RLVR Evaluation Suite

The eight benchmarks reported in the paper, converted to the
verl RLHF parquet schema so they can be
scored directly by verl.trainer.main_ppo in val_only mode. The matching
training data is in the companion LLM-Fusion-Train dataset.

Subsets

Each benchmark is a separate config holding a single test split. They cannot
be splits of one config because extra_info differs per benchmark.

Every row carries an agent_name column that tells verl's rollout which agent
loop to use: tool_agent for BFCL_v3, whose tasks are multi-turn tool calling,
and single_turn_agent for the other seven.

LCB_v5 and LCB_v6 are large for their row counts because the unit tests are
embedded in reward_model.ground_truth; scoring them needs a running
SandboxFusion server.

Upstream benchmarks

Only the prompts and grading targets are redistributed here, in verl's schema.
Refer to each benchmark's own terms of use.

Usage

from datasets import load_dataset

aime = load_dataset("<repo-id>", "AIME2025", split="test")