ngqtrung/mopd-data

MOPD training data (SFT v4 + RL v2, full media)

The data that trains the MOPD domain teachers on Qwen3.5-9B: one SFT set and one RL prompt set per domain (imgr image reasoning, imgp image perception, vid video).

Split	Rows	Format	Media
sft_v4/imgr	1,022,304	chat jsonl (messages, meta)	images in media/
sft_v4/imgp	1,456,403	chat jsonl	images in media/
sft_v4/vid	710,735	chat jsonl	videos in media/
rl_v2/imgr	64,307	verl parquet (prompt, images, reward_model.ground_truth, extra_info)	images embedded (bytes)
rl_v2/imgp	67,915	verl parquet	images embedded (bytes)
rl_v2/vid	60,805	verl parquet	videos in media/

SFT rows: single-turn multimodal questions pooled from public datasets; the assistant turn is a reasoning trace (<think>...</think> + answer, final answer in \boxed{}) written by Qwen3.8-Flash-Next with the gold answer given, then filtered by agreement with the gold. meta keeps the source card, task type and generator. RL rows: prompts with a rule-checkable gold answer (reward_model.ground_truth, extra_info.task_type), decontaminated against our eval benchmarks.

Media

65 store-only zip shards in media_zips/ (523 GB). Unzip all at the repo root:

huggingface-cli download ngqtrung/mopd-data --repo-type dataset --local-dir mopd-data
cd mopd-data && for z in media_zips/*.zip; do unzip -q -n "$z"; done   # -> media/<group>/...

Every media path in the rows is relative to the repo root (media/...). media_index.tsv lists path, bytes and shard. Media processing used in training: images 65,536-1,605,632 px; videos 1 fps, 4-150 frames, ≤ 65,536 px per frame.

Sources and license

The questions and media come from public datasets (e.g. LLaVA-Video-178K, FineVideo, Vript, VideoRFT, LLaVA-OneVision, and others listed in meta.card_id). Each source keeps its own license; check it before use. The reasoning traces are model-generated. Released for research.