S1-Omni-Corpus-10K

<div align="center"> <h1>S1-Omni-Corpus-10K</h1> **An open-source scientific multimodal reasoning dataset subset for S1-Omni** [](https://huggingface.co/datasets/ScienceOne-AI/S1-Omni-Corpus-10K) [](https://huggingface.co/datasets/ScienceOne-AI/S1-Omni-Corpus-10K) [](https://modelscope.cn/datasets/ScienceOne-AI/S1-Omni-Corpus-10K) [](代码仓库/S1-Omni-pro/LICENSE) </div> ## 🧬 Model Introduction **[S1-Omni](https://huggingface.co/ScienceOne-AI/S1-Omni)** is a unified scientific multimodal reasoning model for scientific understanding, prediction, and generation. It is developed by the ScienceOne AI team of the Chinese Academy of Sciences. S1-Omni addresses fragmented scientific AI capabilities with a shared backbone for cross-disciplinary, cross-modal, and cross-task understanding and reasoning, plus science-specific decoders for verifiable outputs. Unified encoding maps natural-language instructions and typed scientific objects, including material CIFs, chemical SMILES, protein sequences, spectra, and scientific images, into shared task representations. Knowledge alignment integrates scientific laws, experimental facts, and expert knowledge into data construction, validation, and training, grounding judgments in evidence. Task-oriented decoding converts the representations into verifiable outputs through specialized decoders for property prediction, spectrum-to-structure reconstruction, protein site and structure prediction, and scientific image generation and editing. <div align="center"> <img src="assets/s1-omni-architecture.png" alt="S1-Omni model architecture" width="95%"> <br> </div> ## 📚 Dataset Introduction **S1-Omni-Corpus-10K** is a representative open-source dataset subset curated from the S1-Omni training corpus. It contains **10,468** complete training samples and is intended for data-pipeline analysis, task-protocol research, and community reproduction. **S1-Omni-Corpus** is a large-scale training corpus for unified scientific multimodal reasoning. It is organized around heterogeneous scientific data unification, expert-experience-aligned reasoning, and domain-native supervision. The corpus covers mathematics, physics, chemistry, biology, materials science, medicine, geography, astronomy, and computer science, and includes scientific question answering, literature reasoning, molecular and materials property prediction, protein function and binding-site prediction, protein-structure-related tasks, spectrum-to-molecular-structure prediction, and scientific image generation and editing. <div align="center"> <img src="assets/s1-omni-corpus.png" alt="S1-Omni-Corpus data distribution" width="95%"> <br> </div> The full S1-Omni-Corpus covers over 200 scientific tasks with million-scale reasoning samples. The S1-Omni-Corpus-10K subset released here preserves the complete training-sample format and can be used as a reference for understanding the S1-Omni data organization and task protocol. ## 🤗 Data Access The open-source dataset is available from: | Platform | URL | | --- | --- | | Hugging Face | `https://huggingface.co/datasets/ScienceOne-AI/S1-Omni-Corpus-10K` | | ModelScope | `https://modelscope.cn/datasets/ScienceOne-AI/S1-Omni-Corpus-10K` | Local files: | File | Description | | --- | --- | | `S1-Omni-Corpus-10K.jsonl` | S1-Omni-Corpus-10K data file. Each line is one JSON training sample. | | `assets/` | Resource directory referenced by multimodal samples such as images and spectra. | ## 🧾 Data Format Each JSONL record contains `data_id`, `messages`, `images`, and `meta`: ```json { "data_id": "XXXXXX", "messages": [ { "role": "user", "content": "User question and scientific-object context" }, { "role": "assistant", "content": "<think>...reasoning process...</think>\n\nModel response and task-specific token" } ], "images": [], "files": [], "meta": { "subject": "...", "task_type": "...", "language": "en", "turns": 1, "label": null } } ``` Field descriptions: - `data_id`: Unique data ID. - `messages`: Data content. `role` contains `user` and `assistant`; `content` contains the user prompt, scientific-object context, and the assistant-side reasoning process and task answer. - `images`: Relative paths of user-input images. Missing values are represented as `[]`. Image and spectra subsets use `assets/...` paths, while text-only records use `[]`. - `files`: Relative paths of user-input files. Only the spectra subset uses non-empty values, pointing to `assets/spectra_jdx/xxx.jdx`; all other records use `[]`. - `meta`: Metadata object containing only `subject`, `task_type`, `language`, `turns`, and `label`. ## 🏷️ Meta Fields | Field | Type / Values | Description | | --- | --- | --- | | `subject` | Discipline | `Chemistry`, `Biology`, `Medicine`, `Materials`, `CS`, `Physics`, `Geography`, `Math`, `Astronomy` | | `task_type` | Task type | Examples include `IR spectrum to SMILES prediction`, `Protein PPI binding site prediction`, `Image editing`, etc. | | `language` | Sample language | `en` / `zh` | | `turns` | int | Number of user turns. `1` means single-turn. | | `label` | object / null | Annotation information. When non-null, it contains `type` and `value`; `value` is the decoder output object for the current scientific task. | ## 🧩 Task-Specific Tokens Task-specific tokens: - `<image_edit>` / `<image_gen>`: Image editing and image generation tasks - `<prot_st>` / `<prot_cla>`: Protein structure and protein classification tasks - `<linear_pre>` / `<linear_cla>`: Linear prediction and classification tasks - `<spectra_st>`: Spectrum-to-molecular-structure prediction task ## 📄 License This dataset is released under the Apache License 2.0. ## 📖 Citation If S1-Omni is useful for your research, please cite our technical report. The formal citation format will be updated after the paper or technical report is released. ```bibtex @misc{s1omni2026, title = {S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation}, author = {ScienceOne AI and Wenge AI}, year = {2026} } ```