erickfm/mimic-melee-by-character

---
language:
- en
license: cc0-1.0
tags:
- melee
- smash-bros
- slippi
- imitation-learning
- controller-inputs
- fighting-games
- pytorch
pretty_name: MIMIC Melee (By Character)
size_categories:
- 100K<n<1M
---

# MIMIC Melee — By Character

Per-character split of [erickfm/mimic-melee](https://huggingface.co/datasets/erickfm/mimic-melee).
Each character has its own subdirectory with ready-to-train PyTorch tensor shards,
enabling character-specific imitation-learning models for Super Smash Bros. Melee.

## Source

Built from [erickfm/mimic-melee](https://huggingface.co/datasets/erickfm/mimic-melee)
using `tools/split_by_character.py` in the [MIMIC](https://github.com/erickfm/MIMIC) repo.
Each game is assigned to a character based on the majority `self_character` value across frames.
All preprocessing (normalization, categorical encoding, stick discretization) is inherited
from the source dataset — the data is identical, just partitioned.

## Dataset statistics

| Split | Games | Frames |
|-------|-------|--------|
| Train | 169,575 | 1,631,777,124 |
| Val | 18,841 | 180,971,668 |
| **Total** | **188,416** | **1,812,748,792** |

- **Characters:** 26 (all playable characters with tournament representation)
- **Total size:** ~2.59 TB
- **Shard size:** ~1-2 GB each
- **Val split:** 10% (seed 42)

## Character distribution

| Character | Games | Train Frames | % of Data |
|-----------|-------|-------------|-----------|
| FOX | 47,539 | 443,452,345 | 27.2% |
| FALCO | 32,007 | 291,899,230 | 17.8% |
| MARTH | 25,113 | 244,108,914 | 15.0% |
| CPTFALCON | 22,481 | 206,274,159 | 12.6% |
| SHEIK | 11,879 | 123,885,020 | 7.6% |
| PEACH | 6,137 | 66,668,337 | 4.1% |
| JIGGLYPUFF | 4,613 | 47,421,314 | 2.9% |
| SAMUS | 3,188 | 36,084,503 | 2.2% |
| ICE_CLIMBERS | 2,350 | 27,021,337 | 1.7% |
| GANONDORF | 2,665 | 24,107,666 | 1.5% |
| LUIGI | 2,236 | 22,938,883 | 1.4% |
| PIKACHU | 1,500 | 16,196,379 | 1.0% |
| YOSHI | 1,435 | 15,183,083 | 0.9% |
| DOC | 1,179 | 12,019,765 | 0.8% |
| DK | 960 | 10,328,792 | 0.6% |
| LINK | 706 | 7,530,259 | 0.5% |
| MARIO | 647 | 6,756,811 | 0.4% |
| GAMEANDWATCH | 564 | 5,227,122 | 0.3% |
| ZELDA | 502 | 4,972,844 | 0.3% |
| KIRBY | 366 | 4,021,788 | 0.2% |
| YLINK | 364 | 3,739,326 | 0.2% |
| ROY | 368 | 3,662,011 | 0.2% |
| NESS | 271 | 3,136,275 | 0.2% |
| MEWTWO | 228 | 2,434,987 | 0.2% |
| BOWSER | 166 | 1,678,142 | 0.1% |
| PICHU | 111 | 1,027,832 | 0.1% |

## Directory structure

```
├── characters.json # Index of all characters with game/frame counts
├── FOX/
│ ├── tensor_manifest.json # Shard list and counts for this character
│ ├── norm_stats.json # Per-column normalization (shared across characters)
│ ├── cat_maps.json # Categorical mappings (shared)
│ ├── stick_clusters.json # Stick cluster centers (shared)
│ ├── train_shard_000.pt
│ ├── train_shard_001.pt
│ ├── ...
│ └── val_shard_000.pt
├── FALCO/
│ └── ...
└── ...
```

## Shard format

Same as [erickfm/mimic-melee](https://huggingface.co/datasets/erickfm/mimic-melee).
Each `.pt` file contains:

```python
{
"states": {feature_name: Tensor}, # normalized game-state features
"targets": {head_name: Tensor}, # controller-input targets
"offsets": [int, ...], # game boundary indices
"n_games": int,
}
```

## Usage

Download a single character:

```python
from huggingface_hub import snapshot_download
snapshot_download(
"erickfm/mimic-melee-by-character",
repo_type="dataset",
local_dir="data/fox",
allow_patterns=["FOX/*"],
)
```

Then train:

```bash
python train.py --data-dir data/fox/FOX --model medium --seq-len 60
```

## Related

- [erickfm/mimic-melee](https://huggingface.co/datasets/erickfm/mimic-melee) — Full dataset (all characters combined)
- [erickfm/mimic-melee-subset](https://huggingface.co/datasets/erickfm/mimic-melee-subset) — Small subset for quick experiments
- [MIMIC](https://github.com/erickfm/MIMIC) — Imitation-learning bot trained on this data
- [Slippi](https://slippi.gg/) — Melee netplay client

## License

CC0 1.0 — Public domain.