Datasets:

MMOPD
/

domain-mopd-data

language:
  - en
task_categories:
  - text-generation
tags:
  - mmopd
configs:
  - config_name: medical
    data_files: data/medical.jsonl
  - config_name: law
    data_files: data/law.jsonl
  - config_name: finance
    data_files: data/finance.jsonl
  - config_name: if
    data_files: data/if.jsonl
  - config_name: tau2
    data_files: data/tau2.jsonl

domain-mopd-data

Prompts of the five domains for multi-teacher on-policy distillation (MOPD and IM-MOPD) of No Pain, More Gain: Iterative Merging for Effective Multi-Teacher On-Policy Distillation (code).

Subset	Rows	Content
`medical`	8,384	four- and five-option MedQA training prompts
`law`	31,654	CaseHOLD and bar-exam prompts; the 91 prompts longer than 15,872 tokens are skipped during training, leaving 31,563
`finance`	7,565	the Finance RL prompts (MMOPD/domain-rl-data)
`if`	11,728	the IF RL prompts (MMOPD/domain-rl-data)
`tau2`	31,718	dialogue contexts from the retail, airline and telecom portions of `inclusionAI/AReaL-tau2-data`, each the prompt for the next assistant turn

Format

One JSON object per line.

medical, law, finance, if: input = user prompt, output = reference answer (answer letter / number; empty for if), meta = source information (for if: the verifiable instructions instruction_id_list and kwargs).
tau2: prompt_text = the dialogue context rendered with the Qwen3 chat template (tool schemas included), ending with the generation prompt of the next assistant turn; sub_domain = retail / airline / telecom.

Sources

MedQA, CaseHOLD, reglab/barexam_qa, FinQA, TAT-QA, nvidia/Nemotron-RL-Ultra-Training-Blends, inclusionAI/AReaL-tau2-data. Please follow the licenses of the source datasets.

Citation

@article{kim2026immopd,
  title   = {No Pain, More Gain: Iterative Merging for Effective Multi-Teacher On-Policy Distillation},
  author  = {Kim, Seonghyeon and Jang, Chaeyun and Lee, Noah and Kim, Boseop and Lee, Juho},
  journal = {arXiv preprint arXiv:2609.34745},
  year    = {2026}
}