MMIF-DeepSeekV41Flash-GLM53Flash-SFT-46K

MMIF-DeepSeekV41Flash-GLM53Flash-SFT-46K is a curated multimodal instruction-following SFT release built from MM-IFEngine prompts and newly generated, conflict-checked tasks. It contains 24,296 answer-only records and 21,869 records with reasoning traces. The retained answers were generated by the Flash variants of DeepSeek V4.1 and GLM 5.3.

This is a mixed-generator release. DeepSeek-V4.1-Flash wrote 37,621 retained records (81.5%), while GLM-5.3-Flash wrote 8,544 (18.5%). Each record contains one selected answer from one generator; the model names in the title do not mean that both models produced every answer.

This is a data release. We have not trained a model on it or measured downstream gains yet. The reported checks describe data curation and release integrity, not model performance.

What's included

Config	DeepSeek-V4.1-Flash	GLM-5.3-Flash	Total
answer_only	19,865	4,431	24,296
thinking	17,756	4,113	21,869
Total	37,621	8,544	46,165

Source stage	DeepSeek-V4.1-Flash	GLM-5.3-Flash	Total
Original MMIF prompts	37,246	0	37,246
New prompts	375	8,544	8,919

The original-prompt portion reruns MMIF examples with several response modes and selects the strongest retained candidate for each export. The new-prompt portion starts from 5,000 generated candidates. Review retained 4,841 tasks across eight constraint profiles and four image domains; 159 candidates were dropped.

Each record includes its source stage, source index, selected rollout variant, and sample ID. Separate provenance files contain request settings, judge decisions, mechanical checks, and a SHA-256 hash of the original training row.

Generation and selection

We generated candidates through OpenAI-compatible and Anthropic-compatible APIs using:

• DeepSeek-V4.1-Flash, used for the original-prompt rollouts and part of the thinking recovery;
• zai-org/GLM-5.3-Flash, used for most new-prompt answers and for parts of review and ranking.

Provider-specific aliases were normalized to the underlying model names shown above. Public provenance omits routing metadata, private endpoints, credentials, and opaque request identifiers.

For each prompt, the pipeline attempted five response modes: native no-thinking, answer-only, and low, high, or max reasoning. It then applied deterministic constraint checks and model-based review. Retained responses had to pass the mechanical checks and meet the configured judge thresholds for visual correctness, instruction following, writing quality, and, where applicable, reasoning quality. The two final export audits found no release-format, image-path, hash, or recorded constraint-check issues.

The new-task pipeline also checked prompts for detectable conflicts among their constraints before rollout. Across the 4,841 retained new tasks, 18,770 deterministic source checks recorded no hard failures, soft failures, or detected prompt conflicts.

Important limitations

• No downstream training or evaluation has been run on this release, so its effect on model quality is unknown.
• The visual and semantic review relies on model judges rather than human gold labels. A mechanically clean record can still contain a subtle visual or reasoning error.
• The new-task native no-thinking candidates did not pass the strict release filter. The 4,577 new-task rows in answer_only were selected across all response modes and had reasoning removed; they should not be treated as outputs generated natively with thinking disabled.
• Reasoning traces may contain scratch work, revisions, or explicit constraint counting. Users should inspect the thinking config and choose a formatting policy that matches their training setup.

Images

To avoid duplicating several gigabytes of image files, this repository stores image paths relative to ChrisDing1105/MMIF-23k. Download and extract the original MMIF-23k image archives, then resolve each value in the images field against that image root.

Loading the data

from datasets import load_dataset

answer_only = load_dataset(
    "ChrisDing1105/MMIF-DeepSeekV41Flash-GLM53Flash-SFT-46K",
    "answer_only",
    split="train",
)

thinking = load_dataset(
    "ChrisDing1105/MMIF-DeepSeekV41Flash-GLM53Flash-SFT-46K",
    "thinking",
    split="train",
)

Each row has the following shape:

{
  "messages": [
    {"role": "user", "content": "<image>\n..."},
    {"role": "assistant", "content": "..."}
  ],
  "images": ["relative/path/to/image.jpg"],
  "sample_id": "...",
  "source_stage": "task1_original_prompts",
  "source_index": 0,
  "rollout_variant": "think_high",
  "generator_model": "DeepSeek-V4.1-Flash"
}

The answer_only assistant content excludes reasoning traces. The thinking assistant content uses <think>...</think> tags followed by the final answer.

Provenance

The files under provenance/ align row-for-row with their corresponding data files. training_row_sha256 hashes the source export row before the five traceability fields shown above were added. Public provenance contains normalized generator and judge model names, generation settings relevant to reproducibility, structured quality scores, and mechanical checks. Generator and judge fields describe different roles: the generator wrote the retained response, while the judge scored or selected it. Routing metadata, private endpoints, credentials, provider-specific aliases, and opaque request identifiers are excluded.

Citation

If you use this release, please cite the original MM-IFEngine paper:

@article{ding2025mm,
  title={MM-IFEngine: Towards Multimodal Instruction Following},
  author={Ding, Shengyuan and Wu, Shenxi and Zhao, Xiangyu and Zang, Yuhang and Duan, Haodong and Dong, Xiaoyi and Zhang, Pan and Cao, Yuhang and Lin, Dahua and Wang, Jiaqi},
  journal={arXiv preprint arXiv:2504.07957},
  year={2025}
}