Datasets:

MMOPD
/

domain-sft-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: tau2
    data_files: data/tau2.jsonl

domain-sft-data

SFT data of the three SFT domain teachers (Medical, Law, Tool Use) of No Pain, More Gain: Iterative Merging for Effective Multi-Teacher On-Policy Distillation (code).

Subset	Rows	Content
`medical`	40,853	correct, deduplicated samples generated by Qwen3.6-35B-A3B from MedQA training prompts (MMOPD/Qwen3-4B-OT3-medical)
`law`	56,123	correct, deduplicated samples generated by Qwen3.6-35B-A3B from CaseHOLD and bar-exam prompts (MMOPD/Qwen3-4B-OT3-law)
`tau2`	32,758	retail, airline and telecom samples from `inclusionAI/AReaL-tau2-data`, after removing overlength samples and telecom task-ID overlap (MMOPD/Qwen3-4B-OT3-tau2)

Format

One JSON object per line.

medical, law: messages = [user, assistant]; the assistant turn is <think> {reasoning}\n</think>\n\n{answer} (OpenThoughts3 thinking opening). meta = prompt id, source and selection statistics.
tau2: prompt_text = the dialogue context rendered with the Qwen3 chat template and the sub-domain tool schemas; completion = the target assistant turn (same thinking format; supervision covers this turn only, earlier thinking is removed from the context); sub_domain = retail / airline / telecom.

Sources

MedQA, CaseHOLD, reglab/barexam_qa, 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}
}