datatally
DataTally：面向 DeepSeek Harness 的可验证数据资产公开使用概况。search_assets / get_asset_profile / compare_assets。DataTally 记录，由你判断。
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:DAAMAAO/datatally#7b4b45ad45fa9ed8097264a6ad047a6bc6f9ed3aDatatally 以 datatally 发布，当前版本为 0.1.5。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。
Official domain: https://datatally.xyz
Help AI find the most valuable data. 帮助 AI 找到最值钱的数据。
DataTally records. AI judges.
For AI, "valuable" means worth the compute — the cost of using data is time and tokens, not money. DataTally helps AI find data worth using.
DataTally is a DeepSeek Harness plugin that aggregates public usage signals of data assets — downloads, citations, stars — into a single, verifiable profile.
Data has no intrinsic properties. A byte count says nothing about value. The only real signal of a data asset's value is how it has been used — by whom, how often, and with what results. DataTally collects those usage signals from public sources and presents them in one place:
| Tool | What it does |
|---|---|
search_assets | Search data asset profiles by keyword or domain |
get_asset_profile | Get the full usage profile of one asset: per-source metrics, timeline, citations |
compare_assets | Compare two assets across dimensions (within the same signal layer) |
Every metric carries its source and fetch time. Nothing is interpreted. Nothing is ranked. DataTally records; you judge.
DataTally never: computes value, weighs signals, ranks assets, recommends "which is better", sells data, or fabricates numbers. The record is DataTally's. The interpretation is the user's.
Layers describe behavior versus interest — not a strength scale. Comparisons are only made within the same layer. A star is not a download. Missing provenance is marked explicitly — a single-source asset says so in its profile summary; a missing license surfaces as null, never a fabricated default.
Requires DeepSeek Harness installed (dsh on your PATH) and Node.js ≥ 22.19.
dsh plugin --profile web add datatally
dsh web
git clone https://github.com/DAAMAAO/datatally.git
cd datatally-plugin
npm install
npm run typecheck
npm test
Load it into a running dsh for end-to-end testing:
dsh web --patch ./cordis.patch.yml
# (replace the `name` in cordis.patch.yml with the absolute path to src/index.ts)
Ask your agent:
"Find datasets for sentiment analysis and compare the two most used ones."
The agent will call search_assets, then compare_assets, and answer with sourced usage data. An actual session transcript:
User：帮我找情感分析的数据集，然后对比两个使用最多的
Agent（调 search_assets("sentiment")）：
- imdb (stanfordnlp/imdb) | domain: nlp | deep: downloads 195,669 | shallow: likes 725 | fetched_at: 2026-09-08T14:35:50Z | snapshot: v2
- glue (nyu-mll/glue) | deep: downloads 826,859 … （共 8 个命中）
Agent（调 compare_assets("nyu-mll/glue", "stanfordnlp/imdb")）：
- downloads [deep]: 791,429 vs 197,595 | same_source: true | same_access: true
- forks [deep]: 178 vs 0 | same_source: false | same_access: true
- likes [shallow]: 741 vs 722 | same_source: true | same_access: true note: cross-source metrics are not directly comparable
Agent 的最终回答只陈述记录（引用次数、来源、抓取时间），不下"哪个更好"的结论——判断留给用户。
The plugin ships with defaults; override per row in your profile patch:
| Field | Default | Meaning |
|---|---|---|
snapshotPath | ./data/snapshot_v2.json | Local snapshot file. Relative paths resolve against the process cwd first, then against the package's bundled data/ seed. |
maxResults | 20 | Search result cap (the limit parameter is clamped to it). |
enableCitationFields | true | Include the citations field in profiles. |
DataTally reads usage data from a local, self-owned JSON file (version 2). Replace the seed by pointing snapshotPath at your own file:
{
"version": "2",
"generated_at": "2026-09-06T12:00:00Z",
"assets": [
{
"asset_id": "stanfordnlp/imdb",
"id_type": "hf",
"name": "imdb",
"domain": "nlp",
"access": "open",
"license": "other",
"verification": "public_api",
"sources": [
{
"source": "huggingface",
"metrics": {
"downloads": { "value": 191564, "signal_type": "deep" },
"likes": { "value": 709, "signal_type": "shallow" }
},
"fetched_at": "2026-09-08T12:32:11.463Z"
}
],
"timeline": [],
"citations": []
}
]
}
Schema principles: structured from day one · every field has a source · evidence attributes are facts, not interpretations · snapshots are self-owned · machine-readable first. Unknown fields (including the legacy asset_class) are rejected loudly; v1 snapshots are rejected with a migration pointer.
The same core, in a terminal (the thin-wrapper form over the plugin core):
datatally profile stanfordnlp/imdb
datatally search sentiment --domain nlp --limit 5
datatally compare HuggingFaceFW/fineweb allenai/c4
datatally export nyu-mll/glue # AI-BOM fact entry: pure facts, no conclusions
datatally catalog # per-sector distributions (multi-source rate, model-use density)
# snapshot location: --snapshot <path> or DATATALLY_SNAPSHOT env
The CLI is read-side only: every command reads a snapshot file. Rebuilding a snapshot (the four-source refresh pipeline) is a maintainer operation in the private workspace; the pipeline never ships in this package or the public repository.
src/
├── index.ts # plugin entry: Config + apply + tool registration
├── core/ # pure query logic (no harness deps) + text rendering
├── tools/ # one defineTool per file: schema + execute + render + UI cards
├── snapshot/ # loader (strict validation, fail-loud), types, AssetId brand
├── schema/ # strict snapshot validator
cli/main.ts # read-side CLI (profile/search/compare/export/catalog)
test/ # read-side tests: 42 unit + keyless drive + goldens
data/snapshot_v2.json # seed snapshot (126 real datasets, 9 sectors)
cordis.patch.yml # bundle patch (installed) / dev patch (checkout)
The maintainer's local checkout additionally holds the private refresh pipeline (src/snapshot/fetchers/ and its tests): excluded from this repository by .gitignore plus a hard deny-gate in the push script, and never present in npm tarballs (the files allowlist has no matching glob).
npm run typecheck — strict TypeScript, no errorsnpm test — builds, then runs the read-side suite (42 tests) including keyless drive tests through the real ctx.tools.execute pipeline@deepseek-ai/cordis, dsh-tools, dsh-llm) are provided by the DeepSeek Harness deployment, exactly like the official dsh tool plugins.The seed snapshot carries real public usage data for 126 open Hugging Face datasets across 9 industry sectors — core (famous benchmarks + sentiment classics), protein, agri-commodity, auto-sales, aviation, insurance-weather, carbon, power, and logistics — aggregated from up to four public sources (fetched 2026-09-08). Each asset carries an optional sector label, searchable like any other field:
| Source | Signals |
|---|---|
| Hugging Face Hub | downloads, likes, model uses (deep / shallow / deep) |
| ModelScope | downloads, likes (mirrored datasets, probed by short name) |
| DataCite | citation counts (when the dataset card carries a DOI) |
| GitHub | stars (shallow), forks/commits (deep) — only for curated dataset→repo mappings whose repo is the dataset's canonical release home (see data/curated.json) |
Provenance discipline: every metric carries its source + fetched_at; model_uses is exact below the scan cap and recorded as model_uses_min (an honest lower bound) at the cap; single-source assets are marked explicitly ("single source only — multi-source aggregation not met"); missing provenance is never fabricated. Hugging Face numbers were fetched through the hf-mirror.com mirror (counts are the mirror's index, which can differ from hf.co's counters); the maintainer-private refresh reads DATATALLY_HF_BASE (default https://huggingface.co) to switch channels.
On calibers, stated neutrally: a platform's own statistics are one caliber among several. Each caliber is recorded separately with its source and fetch time, and a profile that aggregates several calibers is more complete than one that does not — no platform's statistics are attacked or preferred; the record simply states what each caliber shows.
The seed snapshot is produced by the maintainer's private four-source refresh pipeline. The public package and repository are read-side only: they ship and read snapshots, they never fetch. To get a fresh snapshot, ask the maintainer to run the private pipeline and ship the updated data/snapshot_v2.json.
MIT
Issues and pull requests welcome. Please keep contributions within the stated scope: recording usage signals, not interpreting them.