---
task_categories:
- text-generation
tags:
- distillation
- llm
- on-policy-distillation
---

# Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?

This dataset supports the paper [Do We Really Need KL Divergence for On-Policy Distillation of Large Language Models?](https://huggingface.co/papers/2609.33791). It contains training and evaluation data for on-policy distillation (OPD) experiments, including BinaryOPD, positive/negative reward experiments, and consensus multi-teacher on-policy distillation (C-MOPD).

Code and additional details are available in the [GitHub repository](https://github.com/LeapLabTHU/KL-Free-OPD).

## Dataset Structure

| Split    | File                                    | Content                                                        |
|----------|-----------------------------------------|----------------------------------------------------------------|
| `train/` | `DAPO-Math-17k.parquet`                 | DAPO math training set                                         |
| `train/` | `DeepMath-103K-filtered-level6.parquet` | DeepMath, difficulty ≥ 6                                       |
| `train/` | `Eurus-RL-Data-code.parquet`            | Eurus code training set                                        |
| `train/` | `math_and_code.parquet`                 | Balanced DeepMath + Eurus mixture for MOPD/C-MOPD              |
| `eval/`  | `math_eval.parquet`                     | AIME24/25, AMC, MATH500, Minerva, OlympiadBench, IMO-Bench     |
| `eval/`  | `livecodebench_v6.parquet`              | LiveCodeBench v6 evaluation slice                              |
| `eval/`  | `HumanEval.parquet`                     | HumanEval                                                      |
| `eval/`  | `MBPP.parquet`                          | MBPP                                                           |

## Usage

Please refer to the [GitHub README](https://github.com/LeapLabTHU/KL-Free-OPD) for training scripts, reward variants, multi-teacher distillation experiments, and evaluation instructions.