---
title: TacSIm Online Benchmark
emoji: ⚽
colorFrom: blue
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
---

# TacSIm Full Test-Split Online Benchmark

This ZeroGPU Space evaluates one generated CSV for every trajectory in
the complete backend Test split.

## Evaluation workflow

```text
Backend:
Test/
├── 00001.csv
├── 00002.csv
└── ...

User upload:
submission.zip
├── 00001.csv
├── 00002.csv
└── ...
```

The evaluator:

1. securely extracts the submission ZIP;
2. matches CSV files by unique basename;
3. rejects missing or unexpected files;
4. evaluates every matched trajectory;
5. averages results across the entire Test split;
6. produces detailed and aggregate reports;
7. allows the evaluated result to be submitted to the leaderboard.

## Preferred hidden-test setup

Store the official Test split in a **private Hugging Face Dataset**
repository. Configure the Space variable:

```text
TEST_DATASET_REPO_ID=YOUR_USERNAME/tacsim-private-test
```

Optional variables:

```text
TEST_DATASET_SUBDIR=Test
TEST_DATASET_REVISION=main
```

Add a Space secret with read permission:

```text
TEST_DATASET_TOKEN=hf_...
```

The private dataset repository should contain:

```text
Test/
├── 00001.csv
├── 00002.csv
└── ...
```

## Local Test-folder setup

When `TEST_DATASET_REPO_ID` is not configured, the Space reads:

```text
Test/**/*.csv
```

from its own repository.

Do not use this mode in a public Space when the Test CSVs are supposed
to remain hidden.

## Leaderboard setup

Create a separate Dataset repository, then configure:

```text
LEADERBOARD_REPO_ID=YOUR_USERNAME/tacsim-leaderboard
```

Add the write token as a Space secret:

```text
LEADERBOARD_HF_TOKEN=hf_...
```

For compatibility, one shared `HF_TOKEN` may be used instead, provided
it has both private-test read permission and leaderboard write
permission.

## Official benchmark settings

```text
Benchmark version: v2.0-batch
Frame interval: 0.1 seconds
Horizons: 3s, 5s, 10s
Grids: 10×6, 15×10, 20×12, 30×20, 105×68
```

The TacSIm score for each file, horizon, and grid is:

```text
(Spatial Occupancy Similarity + Movement Vector Similarity) / 2
```

The leaderboard Overall score is the arithmetic mean over all test
files, horizons, and grids. Scores are displayed on a 0–100 scale.

## Files uploaded to the Space

```text
app.py
football_trajectory_similarity.py
requirements.txt
README.md
```

When using local Test-folder mode, also upload:

```text
Test/
└── all official test CSV files
```