RLLab/MRRL-Mixed

<!-- mixed-base:start -->

MRRL-Mixed base

base/train is the shared, unbalanced prompt pool for model-specific sampling and calibration.

Sources

• —agentica-org/DeepScaleR-Preview-Dataset, revision b6ae8c60f5c1f2b594e2140b91c49c9ad0949e29.
• —PRIME-RL/Eurus-2-RL-Data, revision 9776b13264b5aaa0b16495fcf086a0a8d86fd655.
• —allenai/Dolci-RL-Zero-Code-7B, revision 054c0a1f9c5fd52566ad9124aa2e702924803cd3.
• —open-r1/verifiable-coding-problems-python_decontaminated-tested, revision 77ace849f7cc8f10d23945c748f9ff464427f83b.
• —sojuL/RubricHub_v1, revision 3837d55971473a872e84879c88f708b8da3ec2ef. Math uses DeepScaleR; Code uses Eurus, Dolci and Open-R1's tested Python problems; IF/Writing use RubricHub's rurbichub_v1_Instruction_Following, rurbichub_v1_Writing and rurbichub_v1_Chat files. Rows retain source repository, revision and subset.

Preparation

• —Validate prompts, Math gold parsing, rubric weights and executable IF rules.
• —Normalize Code interfaces/tests; check available Dolci and required Open-R1 reference programs with the training executor (5s per-test timeout, 120s total execution budget). Keep at most 25 uniformly sampled, prompt-seeded tests per question.
• —Render every Code prompt as ### Question / ### Format from the bare statement (chat, Eurus and Open-R1 wrappers stripped for all sources); drop questions with fewer than 15 words before their first input/example block.
• —Deduplicate normalized prompts across all four tasks; remove exact evaluation-question matches and embedded copies of evaluation questions >=80 characters. This does not detect semantic paraphrases.
• —Require the complete chat prompt to be <=2048 tokens under all three tokenizers: Qwen/Qwen3-4B-Base, RLLab/olmo-3-7b-it-sft, RLLab/gemma-3-4b-text-it.
• —Store outer reward weights separately from rubric-criterion weights; length/RM references are filled by the model-specific sampler.

Overlap sources (upstream, not local evaluation files): Eurus, eval-set/aime24, eval-set/aime25, eval-set/amc23, eval-set/minervamath, eval-set/olympiadbench, eval-set/math500, code_generation_lite/v6, MMLU-Pro/default, ZebraLogic/grid_mode, humanevalplus, mbppplus, IFEval, IFBench, WritingBench, WritingBench-validation, Arena-Hard-v2, GPQA-Diamond, ResearchQA, Creative-Writing-v3.

Task counts: {"code": 35636, "if": 94108, "math": 38685, "writing": 25629}. Refer to the source repositories for dataset licenses.

<!-- mixed-base:end -->