Skip to content

J7Scope v0.2.0rc1 · Research Preview

Pre-release
Pre-release

Choose a tag to compare

@arthurpanhku arthurpanhku released this 28 Oct 06:09
b200432

J7Scope is now a self-hosted observatory for LLM internal representations and reproducible J-lens experiments. Bilingual research is its first application; researchers can run supported weights locally or on their own cloud GPU.

This research preview includes:

  • Installable CLI and bundled viewer/vocabulary, wheel and source distributions, JSON configuration, and Docker/Compose examples.
  • Authenticated, expiring research sessions with separate events, backlog, and recordings; browser connection via HttpOnly cookies.
  • Bounded FIFO model admission shared by generation and patching, connection/request limits, cooperative generation deadlines, and atomic trace publication.
  • Backend/model-adapter plugins and configurable fitting corpora/concept vocabularies.
  • Real pretrained SmolLM2-135M CPU validation, including independent activation capture, identity patching, authenticated JSON/SSE generation, and recorded replay.

Validation: 54 tests passed in the clean release tree (58 in the local workspace including pre-existing draft tests); the pretrained model passed 13 checks; clean wheel startup/recording worked without PyTorch; web build passed with zero npm audit vulnerabilities. GitHub CI also built and ran the demo container. The local Dvalin scan did not report VERIFIED; the manual assessment of its PyTorch .eval() findings is recorded in the verification notes.

Install

Download the wheel attached below:

python -m venv .venv
source .venv/bin/activate
pip install ./j7scope-0.2.0rc1-py3-none-any.whl
j7scope --backend mock --record ./traces

Open http://127.0.0.1:8799/live.html. For real models, install the same wheel with the [hf] extra and run j7scope-validate.

See installation and deployment, extension contracts, and validation details.

Preview limits

Text chat only; tool execution is not included. One model/process, designed for individuals and trusted labs. Small-probe lens fitting checks implementation, not semantic quality or bilingual sharing. Qwen GPU performance and the GPU container profile require hardware validation. Commercial APIs without activations cannot supply J-space read-outs. New traces remain preview:true. This is a GitHub prerelease, not a PyPI publication.

中文:本版将项目定位为可自行部署的 LLM 内部表示观测与实验工具,支持本地和云端自托管模型。真实模型验证已完成,读出语义质量、GPU 性能和双语研究结论仍需独立验证。