ops-huggingface
Dry-run-first Hugging Face Hub operations: search models/datasets, inspect repos, list files, create repos, upload, snapshot, and validate/template model & dataset cards. Mutations require --execute; every receipt records an auth check without exposing token values.
来源信息
- 仓库
- grahama1970/agent-stack-public
- 最近来源活动
- 2026年9月24日 15:51
- 检测到的 SKILL.md 语言
- 英语
- 星标
- 0
- 分支
- 0
安装方式
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
检查来源文件
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
文件资源管理器
3 个文件正在显示 SKILL.md
SKILL.md
来源说明 · 只读预览- name
- ops-huggingface
- description
- Dry-run-first Hugging Face Hub operations: search models/datasets, inspect repos, list files, create repos, upload, snapshot, and validate/template model & dataset cards. Mutations require --execute; every receipt records an auth check without exposing token values.
- provides
- ["ops-huggingface"]
- disciplines
- ["data-engineering","observability-operations"]
# /ops-huggingface
Operate the Hugging Face Hub with dry-run safety defaults.
> RECONSTRUCTED 2026-08-12 from the surviving `.pyc` bytecode after the source
> was lost (never tracked in git, no disk copy survived). The reconstruction is
> faithful to the 3.12 disassembly and is now TRACKED so the skill cannot be
> lost again. See `scripts/hf_ops.py` header for provenance.
## Commands
```bash
run.sh whoami # auth check, no token exposure
run.sh search-models "text classification" --limit 20
run.sh search-datasets "cybersecurity CWE" --limit 20
run.sh repo-info OWASP/example --type dataset
run.sh list-files bert-base-uncased --type model
run.sh validate-card ./README.md --type model
run.sh template-card --type dataset --repo me/mydata --output ./README.md
# Mutations are DRY-RUN by default; add --execute to actually touch the Hub:
run.sh create-repo me/newrepo --type dataset # dry-run receipt
run.sh create-repo me/newrepo --type dataset --execute # actually creates
run.sh upload ./file.json --repo me/data --execute
run.sh snapshot me/data --local-dir ./out --execute
```
## Safety contract
- `create-repo`, `upload`, `snapshot` do nothing to the Hub without `--execute`;
without it they emit a receipt with `dry_run: true, executed: false`.
- Every command records `auth_checked` and `auth_source` (`env` /
`huggingface_hub_cache` / `none`) WITHOUT printing token values.
- `search-*` and `repo-info` are read-only.
## Auth
Reads `HF_TOKEN` / `HUGGINGFACE_HUB_TOKEN` from the environment or the
repository `.env` (loaded at startup), or a cached `huggingface-cli login`
token. No token is ever echoed.
## Requirements
`huggingface_hub` in the active environment. If absent, every command fails
closed with a clear `Missing dependency: huggingface_hub` message.
## QRA/code-training dataset scout note (2026-09-14)
For qra-training and code-slop grader preparation, search dataset families in
parallel and inspect candidates with `repo-info` before any snapshot. Useful
query families observed:
- `code preference`
- `rewardbench code`
- `LiveCodeBench`
- `BigCodeBench`
- `HumanEval`
- `MBPP code`
- `APPS code`
- `CodeContests`
- `SWE-bench`
High-priority candidate IDs from the scout: `Vezora/Code-Preference-Pairs`,
`SWE-bench/SWE-bench_Verified`, `livecodebench/code_generation_lite`, and
`bigcode/bigcodebench`. Treat public datasets as comparison/eval baselines; the
enterprise training unit should still be typed QRA/git/CI/human-label receipts,
not raw benchmark rows.
在 GitHub 查看