PhillNet_Recall / docs /HF_PUBLICATION.md
ayjays132's picture
Upload 121 files
3b3b2bd verified
|
Raw History Blame Contribute Delete
3.58 kB
# Hugging Face publication: two ready-to-upload artifacts
## Model repository
Intended ID: `ayjays132/PhillNet-Recall`. This is a custom-library memory-policy
checkpoint, not an AutoModelForCausalLM checkpoint or a hosted text-generation API.
The model card deliberately omits a misleading text-generation pipeline tag.
The source ZIP is the complete model repository. Extract it and enter the
`PhillNet-Recall` directory. A dry run needs no token and makes no network request:
```bash
python scripts/verify_release.py
python scripts/publish_hf.py
```
Authenticate **locally**, not in a chat message, then create a new private repo:
```bash
python -m pip install '.[hub]'
hf auth login
python scripts/publish_hf.py --publish
```
To create it publicly, explicitly pass `--public --publish`. The script refuses
existing repositories and does not change their visibility. It checks the logged-in
personal namespace, manifest paths, file hashes and sizes, then uploads only those
files. Live SQLite stores, credentials and unlisted files are not uploaded.
If a repository already exists, this script stops. Review its existing files,
visibility and revision before using a separately authorized Hub update process.
No automatic overwrite, history purge, or account permission change is performed.
## Public synthetic demo Space
Extract `PhillNet-Recall-Space-1.2.0.zip` beside the model directory. It contains a
Dockerfile, Space metadata, the installable wheel, the UI and its server, plus a
manifest. The app executes four fixed synthetic demonstrations. It accepts no
personal memory input and creates a fresh in-memory database for each request.
```bash
python scripts/publish_hf.py --target space --space-dir ../PhillNet-Recall-Space
python scripts/publish_hf.py --target space --space-dir ../PhillNet-Recall-Space --public --publish
```
Intended Space ID: `ayjays132/PhillNet-Recall-Demo`. The default visibility remains
private unless explicitly changed on creation. Creating a model repo does not
create a running service; the Docker Space is a separate artifact. Neither repo
was published by this release build. The Docker image and live Hub rendering were
not tested. The app's actual Python routes and all four demonstrations were tested.
## Loading the published policy
Install the supplied wheel and Hugging Face Hub support first. After publication:
```python
from phillnet_recall import RecallMemory, RecallPolicy
policy = RecallPolicy.from_pretrained(
'ayjays132/PhillNet-Recall', revision='YOUR_REVIEWED_COMMIT_SHA')
memory = RecallMemory('local.sqlite3', policy=policy, tenant='me')
```
Remote downloads contain data files only; remote Python and native libraries are
not executed. `trust_remote_code=True` is neither needed nor used. Remote download
execution was not verified here. Local `save_pretrained` / `from_pretrained` are
tested. A repository name is not a substitute for pinning and reviewing a revision.
## Post-upload verification
Confirm card images render, compare the repository commit to the uploader receipt,
download the data at that commit into a clean environment, and rerun the local
installation and example. Keep live memory databases out of the model repository.
Use authenticated private storage for production memory. The demo is synthetic,
not a multi-user memory-hosting product or a Hugging Face endorsement.
Official references: https://huggingface.co/docs/hub/model-cards ;
https://huggingface.co/docs/huggingface_hub/guides/upload ;
https://huggingface.co/docs/hub/spaces-sdks-docker . Reviewed 2026-11-11.