poscape/UniWeaver-10K

UniWeaver: Harnessing Unified Multimodal Models for Interleaved Text-Image Generation

Project Page · Code · Benchmark

---

Overview

UniWeaver-10K is a dataset of 10,000 goal-only requests with four-stage planning trajectories, giving 40,000 supervised samples in total. The requests cover 20 task families, 500 each, and split 2,510 easy, 4,980 medium, and 2,510 hard. Supervision is planning only.

Files

UniWeaver-10K/
├── manifest.json
├── manifest.jsonl
├── README.md
└── trajectories/
    ├── uw10k_architecture_design_style_001.json
    ├── uw10k_architecture_design_style_002.json
    └── ... (10,000 files)

manifest.json lists every trajectory with its family, difficulty, and SHA-256. manifest.jsonl carries the same index one record per line, which is what the dataset viewer pages through. trajectories/uw10k_*.json holds task_key and raw_attempts with stages plan1 to plan4. Each stage is a list holding one attempt with system_text, raw_input, and raw_output; the task request sits in raw_attempts.plan1[0].raw_input.

{
  "task_key": "uw10k_architecture_design_style_001",
  "raw_attempts": {
    "plan1": [{
      "system_text": "You construct useful illustrated reports in four stages. ...",
      "raw_input": "Stage 1 — task scope and sequence.\n\nTask request:\nFor a design enthusiast ...",
      "raw_output": "{\"title\": \"Recognizing Art Nouveau Architecture\", \"steps\": [...]}"
    }],
    "plan2": [{"...": "..."}],
    "plan3": [{"...": "..."}],
    "plan4": [{"...": "..."}]
  }
}

Usage

import json
from pathlib import Path

root = Path("UniWeaver-10K")
sample = json.loads((root / "trajectories/uw10k_architecture_design_style_001.json").read_text())
print(sample["task_key"], list(sample["raw_attempts"]))

Train with harness/sft_train/run_sft.sh <trajectories dir> <output dir> from the released code.

Ethics

Requests were written for this dataset and screened for objectionable content. The release contains no personal data and no generated images.

Citation

@article{chen2026uniweaver,
  title   = {UniWeaver: Harnessing Unified Multimodal
             Models for Interleaved Text-Image Generation},
  author  = {Hongxu Chen and Zixuan Zhong and Donghao Zhou
             and Haodong Li and Xiaomin Yu and Tianhao Li
             and Shiyao Yu and Zhongwei Wan and Pheng-Ann Heng},
  year    = {2026},
  note    = {Project page}
}