OpenDCAI/Omni-Edu

Omni-Edu — instruction-tuning mixture

69,999 supervised instruction examples (~158M characters) covering K-12 subject competence, curriculum grounding, diagnostic reasoning, pedagogical action and general-purpose instruction. 12,146 rows (17.4%) are multimodal; every image referenced by the JSONL ships in this repository under images/.

Composition

Format

JSON Lines, ShareGPT style, exactly three top-level keys per row:

{
  "messages": [
    {"role": "system",    "content": "..."},
    {"role": "user",      "content": "Problem:\n<image> ..."},
    {"role": "assistant", "content": "..."}
  ],
  "images": ["images/0a/0af6e8ac....png"],
  "extra_info": "{\"category\": \"subject_competence\", \"dataset\": \"geometry3k-answer-only\", ...}"
}

• —messages — system, then alternating user / assistant. Multi-turn rows go up to 21 messages; 94.5% are single-turn.
• —images — 0 to 16 entries, relative to the repository root; 82.6% of rows are text-only, 16.2% carry one image, and the rest carry between 2 and 16. The <image> marker inside the user turn shows where each image belongs, and 131 rows reference the same image more than once.
• —extra_info — a JSON-encoded string (not an object); json.loads it to reach category, dataset, uid, system_prompt_id, system_prompt_version and, where present, kcenter / source_license / language.

dataset_info.json is the llama-factory registration file for this mixture.

Loading

from datasets import load_dataset

ds = load_dataset("OpenDCAI/Omni-Edu", "OmniEdu", split="train")
print(ds[0]["messages"], ds[0]["images"])

Images resolve against the repository root, so either download the full snapshot or prefix images/... with the repository URL when serving rows.