Dataset card for MMK12-QWEN35-9B-VERIFIED

Strongich/MMK12-QWEN35-9B-VERIFIED is a filtered teacher-trace dataset built from FanqingM/MMK12 using Qwen/Qwen3.5-9B. The goal is to provide high-quality multimodal reasoning traces that can be used for off-policy distillation into smaller models.

Each row keeps the original MMK12 example fields and adds a responses column. The responses column contains a list of verified Qwen responses that survived the filtering pipeline, so a question can end up with fewer than 8 kept responses even though 8 rollouts were sampled initially.

Creation Process

1. Rebuild the MMK12 splits

The original FanqingM/MMK12 test split contains only multiple-choice questions. Because of that, the original train and test splits were merged and then re-split deterministically with:

• test_size = 2000
• seed = 1337

The published train and test splits in this dataset therefore do not match the original MMK12 split boundary.

2. Generate 8 Qwen rollouts per question

For each question-image pair, Qwen/Qwen3.5-9B was sampled for up to 8 rollouts with the following system prompt:

R1_STYLE_SYSTEM_PROMPT_GRPO = (
    "You are a visual agent and should provide concise answers."
    "The user asks a question, and the Assistant solves it. "
    "The assistant first thinks about the reasoning process in the mind "
    "and then provides the user with the answer. "
    "Let's think step by step and output the final answer within \\boxed{}."
)

Rollout generation used a maximum completion budget of 16384 tokens.

3. Keep only questions with strong rollout agreement

After rollout evaluation, only questions for which Qwen produced at least 6 correct rollouts out of 8 were kept. This biases the dataset toward examples where the teacher is consistently correct rather than occasionally lucky. As a result, the retained questions are useful not only for off-policy distillation, but also as a higher-confidence prompt pool for on-policy distillation setups such as GOLD, where stable teacher behavior is desirable.

4. Filter individual responses

Even for kept questions, each candidate response still had to pass several filters:

• it had to contain </think>, so incomplete reasoning traces were dropped
• the combined prompt and response length had to be <= 16384 tokens
• its 4-gram repetition ratio had to be <= 0.30

Intended Use

This dataset is meant for off-policy distillation of smaller models from a stronger multimodal teacher. In practice, the responses list gives multiple filtered teacher traces per question, which makes the dataset useful for SFT-style training, response selection, or other teacher-student training setups.

Code Availability

The dataset generation code is not included in this release. It will be released later.