UPLOAD.md

# AcuBench → Hugging Face: upload runbook

This runbook is for **you** (the author) to run with **your own** HF token.
The packaging step did **not** log in, create accounts, enter tokens, or upload
anything. Everything below is copy-pasteable; replace `<username>` with your HF
username (and adjust repo names if you like).

Two staging trees are ready:

- `acubench/hf/` → the HF **dataset** repo (build-script pattern; no
  AcuKG-derived labels shipped).
- `acubench/hf_space/` → the HF **Space** (Gradio leaderboard).

---

## 0. One-time setup

```bash
pip install -U "huggingface_hub[cli]"
huggingface-cli login          # paste your token (needs write scope)
huggingface-cli whoami         # confirm you are logged in
```

---

## 1. Create + upload the dataset repo

```bash
# Create the dataset repo (idempotent: --exist-ok won't fail if it exists).
huggingface-cli repo create acubench --repo-type dataset --exist-ok

# Upload the ENTIRE staging folder as the repo root.
huggingface-cli upload <username>/acubench \
    /Users/yutaewan/test/fromhope1/merdian/acubench/hf \
    . \
    --repo-type dataset \
    --commit-message "Add AcuBench (build-script pattern: skeleton + build/eval scripts, no AcuKG labels)"
```

**Guardrail double-check before/after upload** — the dataset repo must contain
ONLY these files. It must NOT contain `acubench.jsonl` or any raw AcuKG data:

```bash
ls /Users/yutaewan/test/fromhope1/merdian/acubench/hf
# expected: README.md  NOTICE  LICENSE  UPLOAD.md
#           build_acubench.py  eval.py
#           who_acupoints.csv  meridian_adjacency.csv  sample_labels.jsonl
```

Then view it at `https://huggingface.co/datasets/<username>/acubench`.

---

## 2. Create + upload the Space (Gradio leaderboard)

```bash
# Create a Gradio Space.
huggingface-cli repo create acubench-leaderboard --repo-type space --space_sdk gradio --exist-ok

# Upload the Space staging folder.
huggingface-cli upload <username>/acubench-leaderboard \
    /Users/yutaewan/test/fromhope1/merdian/acubench/hf_space \
    . \
    --repo-type space \
    --commit-message "Add AcuBench leaderboard (operator-held gold)"
```

The Space builds automatically and serves the leaderboard (seeded with the
paper's reference baselines). View it at
`https://huggingface.co/spaces/<username>/acubench-leaderboard`.

---

## 3. Private gold handling (operator-held gold)

The private **test** gold is **never** uploaded to the public Space or dataset
repo. You build it locally and keep it on your machine:

```bash
# You need your own AcuKG clone (its labels are not redistributed):
git clone https://github.com/<acukg-owner>/AcuKG.git /path/to/acukg

# Build the labels + splits locally (byte-reproduces the canonical dataset):
cd /Users/yutaewan/test/fromhope1/merdian/acubench/hf
python3 build_acubench.py --acukg /path/to/acukg --out ~/acubench_gold
# -> ~/acubench_gold/acubench.jsonl  (keep this PRIVATE; it holds the test labels)
```

If you ever want automated scoring *inside* the Space (optional, not required by
the operator-held-gold flow), add the gold as a **private Space secret/file**
rather than committing it:

- **As a file** (private): `huggingface-cli upload <username>/acubench-leaderboard
  ~/acubench_gold/acubench.jsonl gold/acubench.jsonl --repo-type space` — but
  only if the Space is **private**; do NOT do this on a public Space.
- **As a secret**: Space → Settings → *Variables and secrets* → add a secret
  (e.g. `ACUBENCH_GOLD_PATH` or the gold contents) and read it via
  `os.environ` in `app.py`. Recommended only if you make the Space private.

The default, safe design keeps the Space public and the gold entirely on your
machine (see next section).

---

## 4. Operator scoring + appending leaderboard rows

When a participant sends you a predictions JSONL (they do **not** have the
gold), score it locally and commit the resulting row:

```bash
cd /Users/yutaewan/test/fromhope1/merdian/acubench/hf_space

# Score against your private, locally-built test gold:
python3 score_submission.py \
    --pred /path/to/participant_preds.jsonl \
    --gold ~/acubench_gold/acubench.jsonl \
    --who  ../hf/who_acupoints.csv \
    --eval ../hf/eval.py \
    --split test \
    --model "Participant Model v1" \
    --submitter "Participant Name" \
    --out submissions/010_participant_model.json

# Push just the new row to the Space:
huggingface-cli upload <username>/acubench-leaderboard \
    submissions/010_participant_model.json \
    submissions/010_participant_model.json \
    --repo-type space \
    --commit-message "Leaderboard: add Participant Model v1"
```

The Space picks up the new `submissions/*.json` on reload. No gold labels ever
leave your machine.

---

## 5. TODO / decisions for you

- [ ] **Confirm the license.** The packaging step defaulted to **CC BY 4.0**
      for the data/knowledge artifacts (WHO skeleton, adjacency, sample) and
      **MIT** for the code (`build_acubench.py`, `eval.py`). See `LICENSE`.
      If you prefer a single license, edit `LICENSE` and the `license:` field in
      `README.md`'s YAML frontmatter accordingly.
- [ ] **Fill in the AcuKG URL** in `README.md`, `NOTICE`, and the build/upload
      commands above (`<acukg-owner>`), and add the AcuKG citation.
- [ ] **Confirm the citation** block in `README.md` (author list, venue/year).
- [ ] Consider **contacting AcuKG's authors** for an explicit LICENSE or
      redistribution permission (recommended, not blocking).
- [ ] Decide whether the Space stays **public** (default, operator-held gold) or
      goes **private** to enable in-Space automated scoring (§3).