@snow-the/dsh-research-lab
DeepSeek Harness 的研究实验室：AutoSci 风格的维基（实验/文献/决策）、ASI-Bench 级别的评估严谨性记录、arXiv 摘要、对抗性论文审阅和学术写作检查。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:snow-The/dsh-research-lab#56d3e5b2dfe997590f3edeac1611464f290ab127Research Lab 以 @snow-the/dsh-research-lab 发布，当前版本为 0.2.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Research-lab toolkit for DeepSeek Harness: a wiki-centric, eval-rigor-driven research workflow distilled from AutoSci (wiki as the center of a research project), ASI-Bench (evaluation rigor), daily-arXiv-ai-enhanced (digest), ChatPaper (paper review), nature-skills / research-writing-skill (academic writing), Supervisor-Skills (senior-review instinct), and hard lessons from real distillation experiments (hypothesis → verdict loops, eval 口径 discipline).
| tool | purpose | inspired by |
|---|---|---|
rlab_wiki | write wiki pages (experiment/literature/decision/todo) under <project>/.rlab/wiki/ + rebuild index | AutoSci |
rlab_experiment | hypothesis → prediction → evidence → verdict → conclusion pages | AutoSci + falsifiability discipline |
rlab_bench | per-project eval ledger (bench.jsonl); flags 口径 (protocol) mismatches, e.g. mixing test/dev splits | ASI-Bench |
rlab_paper_review | arXiv id → adversarial 8-point review checklist (also applies to your own papers) | ChatPaper + ASI-Bench |
rlab_arxiv_digest | keyword digest of recent arXiv papers → <project>/.rlab/digests/<date>.md | daily-arXiv-ai-enhanced |
rlab_writing | line-anchored academic-expression check (passive/weak verbs/vague quantifiers/filler/term consistency) | nature-skills, research-writing-skill |
rlab_related | self-building keyword retrieval: NLP tokenization + SQLite FTS5, no preset lexicon — terms are TF-IDF-mined from your docs (EN words + ZH n-grams) and expanded iteratively (search → mine → merge → repeat) | hamuleite-style corpus, Rocchio relevance feedback |
rlab_status | one-screen project state: wiki counts, bench report, open TODOs | — |
<project>/.rlab/
wiki/
index.md # auto-rebuilt
experiment/<id>.md # hypothesis → verdict
literature/<id>.md # paper notes (link from digests)
decision/<id>.md # ADRs: why you chose X
todo/<id>.md # open items
related.db # FTS5 corpus + auto-built lexicon (rlab_related)
bench.jsonl # one eval row per line (JSON)
digests/<date>.md # arXiv digests
Every row must state metric, split, and for MTEB tasks the hf_subset. The report
warns whenever a task's history mixes splits — comparing test-vs-dev scores is the classic
way papers fool themselves (and reviewers).
# from ~/.dsh/profiles/web
pnpm add @snow-the/dsh-research-lab@file:../.dsh-starter/plugins/dsh-research-lab
# add "dsh-research-lab" to dsh.profile.bundles, then restart dsh web
pnpm install && pnpm build # tsc declarations + esbuild bundle → dist/