model_cache_code_step8000/HF_UPLOAD.md

# Hugging Face Upload Plan

Use the modern `hf` CLI, not deprecated `huggingface-cli`.

## Recommended Repos

Use two repos:

1. Model/reproducibility repo for this package:
   - checkpoints
   - code snapshot
   - derived cache archive
   - logs and metadata

2. Dataset repo for raw PanTS train/test package:
   - raw split archive
   - PanTS repo README/license
   - split/caption/canonical metadata

## Upload Model Package

```bash
cd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_wan22_b16_9k_step8000_20260523_100252
hf auth whoami
hf upload-large-folder Neuripsused/MA048181 . --type model --num-workers 8
```

For a private upload:

```bash
hf repos create Neuripsused/MA048181 --type model --private --exist-ok
hf upload-large-folder Neuripsused/MA048181 . --type model --num-workers 8
```

## Upload Raw Data Package

```bash
cd /scratch/user/yuhwang/artifacts/twoframe/hf_upload/pants_raw_train_test_20260523_102411
hf repos create Neuripsused/MA048181 --type dataset --private --exist-ok
hf upload-large-folder Neuripsused/MA048181 . --type dataset --num-workers 8
```

## Verify Before Upload

```bash
sha256sum -c SHA256SUMS.txt
find . -maxdepth 3 -type f | sort | sed -n '1,200p'
```

## Suggested Repo README Language

This artifact contains derived training cache and checkpoints. The raw PanTS source data is distributed separately because it is a much larger dataset-layer artifact. The last complete checkpoint is step 8000, with both EMA and non-EMA transformer weights available.