README.md

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

# domain-rl-data

RL (GRPO) prompts of the two RL domain teachers (Finance, IF) of [No Pain, More Gain: Iterative Merging for Effective Multi-Teacher On-Policy Distillation](https://arxiv.org/abs/2609.34745) ([code](https://github.com/shkim0824/immopd)).

| Subset | Rows | Content |
|---|---|---|
| `finance` | 7,565 | numeric-answer prompts from the training splits of FinQA and TAT-QA ([MMOPD/Qwen3-4B-OT3-finance](https://huggingface.co/MMOPD/Qwen3-4B-OT3-finance)) |
| `if` | 11,728 | verifiable instruction-following prompts from the `rlvr1` blend of `nvidia/Nemotron-RL-Ultra-Training-Blends` ([MMOPD/Qwen3-4B-OT3-if](https://huggingface.co/MMOPD/Qwen3-4B-OT3-if)) |

## Format

One JSON object per line.

- `finance`: `input` = question with the report table and text, `output` = numeric answer, `meta` = source and answer scale.
- `if`: `messages` = [user prompt], `instruction_id_list` and `kwargs` = the verifiable instructions and their arguments.

## Sources

[FinQA](https://github.com/czyssrs/FinQA), [TAT-QA](https://github.com/NExTplusplus/TAT-QA), [nvidia/Nemotron-RL-Ultra-Training-Blends](https://huggingface.co/datasets/nvidia/Nemotron-RL-Ultra-Training-Blends). Please follow the licenses of the source datasets.

## Citation

```bibtex
@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}
}
```