Skip to content

Sage-SciPerception v0.1.0 — Initial Public Release

Latest

Choose a tag to compare

@yuhangzang yuhangzang released this 06 Nov 13:24

Sage-SciPerception connects scientific representations to language models through task-specific projectors. This initial public release brings together the first checkpoints, inference tools, and bilingual examples.

Included

  • Nine GLM-5.3-Flash task-specific projectors, covering crystal formation energy, XRD classification, protein interface sites, protein structure ranking, molecular identification from spectra, DNA splice sites, ncRNA families, ECG classification, and EEG sleep staging.
  • Index-based checkpoint downloads and strict CPU loading with matching configuration and weights resolved from one Hugging Face snapshot.
  • A vLLM crystal example that inserts scientific vectors into a fully templated prompt-embedding sequence.
  • English and Chinese documentation, framework animations, and four scientific showcases.

Get started

git clone --branch v0.1.0 --depth 1 https://github.com/InternLM/Sage-SciPerception.git
cd Sage-SciPerception
python -m pip install -r requirements.txt
python load_projector.py glm-5.3-flash/cif/jarvis \
  --revision 7c228fc03394027fed19bda1ff26388394b0031a \
  --local-dir ./checkpoints

Continue with the CPU loading quick start, vLLM example, and results with metric definitions.

Weights and reproducibility

Weights and configurations are available on Hugging Face. This release pairs code tag v0.1.0 with HF snapshot 7c228fc03394027fed19bda1ff26388394b0031a. The example above selects that exact snapshot. Use the language-model snapshot and scientific input representation declared in each checkpoint configuration.

Scope and dependencies

Each checkpoint is trained for its respective task. Scientific feature extraction and task-adapted scientific models are separate dependencies; adapted encoder weights for DNA, MPBind, and ncRNA are not included in this checkpoint release. The crystal client starts from prepared CHGNet atom states. GPU serving requires a compatible vLLM environment and the matching language-model weights.

Checkpoint download and strict CPU loading have been exercised for CIF/JARVIS. The example CLI entrypoints and bilingual media were also checked.

License details have not yet been specified for this release.

中文说明

这是书生·元征的首个公开版本,包含首批 9 个 GLM-5.3-Flash 任务专用 projector、按索引下载与 CPU 加载工具、vLLM 晶体接入示例,以及中英文文档和动图。模型权重与配置固定在上述 HF 快照,科学输入需按照各任务配置准备;部分任务适配科学模型尚未包含在本次发布中。本次发布尚未指定许可证信息。

Citation

@misc{sage_sciperception,
  author       = {{InternLM}},
  title        = {{Sage-SciPerception}},
  version      = {0.1.0},
  note         = {GitHub repository},
  year         = {2026},
  url          = {https://github.com/InternLM/Sage-SciPerception}
}