Datasets:

BytedTsinghua-SIA
/

Open-MOPD-Data

pretty_name: Open-MOPD Data
language:
  - en
task_categories:
  - text-generation
tags:
  - open-mopd
  - math
  - code
  - instruction-following
  - reinforcement-learning
size_categories:
  - 1M<n<10M
configs:
  - config_name: rl_prompt_mix
    data_files: rl_prompt_mix/train.parquet
  - config_name: sft_openr1_math_93k
    data_files: sft/openr1_math_93k/train.parquet
  - config_name: sft_ocr_50k
    data_files: sft/ocr_50k/train-00000.parquet
  - config_name: sft_instruction_nemotron_aligned
    data_files: sft/instruction_nemotron_aligned/*.parquet
  - config_name: eval_aime24
    data_files: eval/math/aime24.parquet
  - config_name: eval_aime25
    data_files: eval/math/aime25.parquet
  - config_name: eval_livecodebench_v5
    data_files: eval/code/livecodebench_v5.parquet
  - config_name: eval_livecodebench_v6
    data_files: eval/code/livecodebench_v6.parquet
  - config_name: eval_ifeval_aligned
    data_files: eval/if/ifeval_aligned.parquet
  - config_name: eval_ifbench_test_aligned
    data_files: eval/if/ifbench_test_aligned.parquet

Open-MOPD Data

This repository contains the training and evaluation data released with Open-MOPD, including mixed-domain supervised fine-tuning data, the shared RL/OPD prompt mixture, and six evaluation benchmarks.

Dataset contents

Configuration	Description	Examples
`rl_prompt_mix`	Shared math, code, and instruction-following prompts for RL and OPD	86,931
`sft_openr1_math_93k`	Math SFT data in a unified think-tag format	93,733
`sft_ocr_50k`	Sampled OpenCodeReasoning data in a unified format	50,000
`sft_instruction_nemotron_aligned`	Instruction-following SFT data in a unified format	820,039
`eval_aime24`, `eval_aime25`	Math evaluation sets	30 + 30
`eval_livecodebench_v5`, `eval_livecodebench_v6`	Code evaluation sets	167 + 175
`eval_ifeval_aligned`, `eval_ifbench_test_aligned`	Instruction-following evaluation sets	541 + 300

The three SFT domains are balanced by response-token count rather than example count. After balancing, math, code, and instruction following contribute approximately 37.3%, 28.1%, and 34.6% of training response tokens.

rl_prompt_mix/manifest.json records the construction and decontamination of the shared prompt mixture. In particular, code training prompts explicitly exclude LiveCodeBench.

Evaluation protocol

Metrics are averaged per dataset, then per domain, followed by a macro-average across the three domains.

Math: AIME24 and AIME25, avg@64, temperature 0.6.
Code: LiveCodeBench v5 and v6, avg@10, temperature 1.0.
Instruction following: IFEval and IFBench_test, n=1, temperature 1.0, with enable_thinking=true.

All evaluations use max_model_len=32768, top_p=0.95, top_k=-1, and stop_token_ids=[128012]. Sampling-related columns preserved in the Parquet files are legacy construction metadata and do not define the final protocol; use the settings documented above.

Related models

BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-Final
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-MixSFT
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-RL-Math
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-RL-Code
BytedTsinghua-SIA/Open-MOPD-SmolLM3-3B-RL-IF

Sources and licensing

The release is derived from OpenR1-Math, OpenCodeReasoning, Instruction-Nemotron, and the listed public evaluation benchmarks. Users must follow the licenses and terms of the corresponding upstream sources when using or redistributing each configuration.