Sovereign 8s.is mirror of 8b-is/hf-cards
- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
Datasets and spaces carry their own cards, so the tool now covers all three kinds: 76 in total. Two of the bugs it hit are worth keeping in the message. A model and a dataset can share a name — ultrawhale-dogfood does — and the first version read the model path for both, which turned a 40 KB dataset card into a 659-byte stub. The kind prefix is not cosmetic. The same mistake was waiting in the upload call, which hardcoded the model type. Dataset cards also nest `configs:` in their front matter, exactly as model cards nest `model-index`. A flat parser rewrites both into something the Hub rejects, so any front matter containing either key is now preserved byte for byte and only the body is managed. One card had no README at all: the quant-sign witness set, 2,627 files. It gets prose rather than a stub, and every claim in it is checkable against a file in the bucket or a named script. |
||
| .gitignore | ||
| align.py | ||
| README.md | ||
hf-cards — every model card, wired in and aligned
One pass over every Hugging Face model under the account, so a reader landing on any of them knows three things immediately: whose family it belongs to, what is actually inside the repository, and where the rest of it lives.
python3 align.py --dry-run # proposals land in ./proposed, nothing is touched
python3 align.py --apply # upload the aligned cards
python3 align.py --apply --only PeetPedro/kompress-v17
What it does, in order of what matters
- Idempotent. The managed content sits between two markers. Re-running replaces that block and leaves everything else alone, so a hand-written paragraph survives any number of passes.
- Nothing invented. "What is inside" is read from the repository's own file
listing; lineage comes from the card's own YAML (
base_model,datasets,pipeline_tag). No benchmark, no number, no claim is generated. - Junk is removed and named. Several cards carried an unrelated
Model API provider: CometAPIadvert committed by a template, complete with an empty## Infrastructureheading. Those sections are dropped and reported, because a card that advertises a third party is not a card about the model. - Nested front matter is not rewritten. Cards with a
model-indexblock keep their YAML byte for byte — a flat parser that "helpfully" re-serialises it produces a card the Hub rejects (model-index[2].results is required).
What a managed block looks like
<!-- 8BIS-CARD-START managed by hf-cards/align.py -->
## quantal-ternary
Part of the **8b.is AI training** family ...
| | |
|---|---|
| repository | `PeetPedro/quantal-ternary` |
| contents | safetensors weights, sharded JSON weight maps, tokenizer |
| files | 176 |
| lineage | **base model** · `Qwen/Qwen2.5-0.5B` · **trained on** · `PeetPedro/ultrawhale-dogfood` |
**Family** · [8b.is](https://8b.is) · [classroom](https://classroom.vaked.dev) · [the library](https://pocoo.vaked.dev) · [github](https://github.com/8b-is)
<!-- 8BIS-CARD-END -->
The run that produced the current cards
37 models · 36 carried the advert template · 37 now carry a managed block and the
family tags 8b-is, sovereign, quantal. Standard library only; talks to the
Hub with the token from HF_TOKEN or ~/.cache/huggingface/token (the file, so it
also works from a launchd daemon that never sees your shell).