Fix Space startup: sync full app source

src/about.py

@@ -1,66 +1,38 @@
 TITLE = """ TSFM Realworld Bench Leaderboard """
 
 INTRODUCTION_TEXT = """
-**TSFM Realworld Bench** evaluates time series foundation models on **live TS-Bench
-data** with **zero-shot API inference** via [TSFM.ai](https://tsfm.ai/).
-
-This Space is **UI-only**: it displays leaderboard results pushed from a local runner
-(or GitHub Actions). Evaluations are **not** run inside the container, so the page loads
-instantly.
+**TSFM Realworld Bench** evaluates time series foundation models on **live Open-Meteo
+weather data** with **zero-shot API inference** via [TSFM.ai](https://tsfm.ai/).
+Scores are normalized against **Chronos-Bolt-Base**. The leaderboard shows
+**metric values** and **ranks** in separate tables (MASE, CRPS, MAE, RMSE, sMAPE,
+MSIS, ND, NRMSE, MAPE).
 """
 
 LLM_BENCHMARKS_TEXT = """
-## Architecture
+## How to participate
 
-| Component | Role |
-|---|---|
-| **HF Space** | Gradio leaderboard UI (reads `results/` CSVs) |
-| **Local runner** | Every hour: TS-Bench collect → TSFM.ai eval → push to HF |
-| **GitHub Actions** | Optional hourly CI eval + upload (same as local runner) |
+1. Clone [TSFM_Bench](https://github.com/zhouziyu02/TSFM_Bench)
+2. Implement your model against the streaming data interface (`tsfm_bench.data`)
+3. Run evaluation and write `results/<MODEL_NAME>/all_results.csv`
+4. Open a pull request; results sync to this leaderboard
 
-### Local runner (recommended)
+## Metrics
 
-```bash
-# .env needs TSFM_API_KEY and HF_TOKEN
-cd /path/to/TSFM_Bench
-source .venv/bin/activate
-python scripts/run_online_daemon.py        # loop: eval + push every 1 h
-python scripts/run_online_daemon.py --once # single cycle
-```
+- **MASE** — Mean Absolute Scaled Error (normalized)
+- **CRPS** — Mean Weighted Sum Quantile Loss (probabilistic forecast quality)
+- **CRPS_Rank** / **MASE_Rank** — average per-dataset rank (lower is better)
 
-Or from `space/`:
+## Data interface
+
+Live weather series are streamed from **Open-Meteo** (`OpenMeteoDataSource`). Each evaluation
+holds out the most recent forecast horizon as ground truth and calls TSFM models through
+the **TSFM.ai** API for zero-shot forecasts.
 
 ```bash
-python scripts/run_online_daemon.py --once
+export TSFM_API_KEY=your_key
+python scripts/run_online_eval.py
+python scripts/run_online_daemon.py --interval-minutes 60
 ```
-
-### HF Space settings
-
-No secrets required for display-only mode. Optionally set:
-
-| Variable | Default | Purpose |
-|---|---|---|
-| `TSFM_LIVE_EVAL` | `0` | Set `1` only if you want eval inside the Space (not recommended) |
-| `TSFM_BENCH_INTERVAL_SECONDS` | `3600` | Countdown interval shown in UI |
-
-### Required for local runner (.env)
-
-| Variable | Purpose |
-|---|---|
-| `TSFM_API_KEY` | TSFM.ai API key |
-| `HF_TOKEN` | Hugging Face write token for pushing results |
-
-## Metrics
-
-- **MASE** — Mean Absolute Scaled Error
-- **CRPS** — Mean Weighted Sum Quantile Loss
-- Ranks are averaged per-dataset on raw metrics (lower is better)
-
-## Aggregation (GIFT-Eval style)
-
-1. Compute evaluation metrics on each dataset (raw values).
-2. Rank models per dataset.
-3. Aggregate with **geometric mean** (metrics) and **arithmetic mean** (ranks).
 """
 
 CITATION_BUTTON_LABEL = "Copy citation"