Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,76 @@
+---
+license: mit
+tags:
+- melee
+- smash-bros
+- imitation-learning
+- game-ai
+- pytorch
+pretty_name: "MIMIC: Melee Imitation Model for Input Cloning"
+---
+
+# MIMIC Checkpoints (No Opponent Inputs)
+
+Trained model checkpoints from the [MIMIC](https://github.com/erickfm/MIMIC) project — an imitation-learning bot that predicts human controller inputs from game state in Super Smash Bros. Melee.
+
+These checkpoints are from the **Phase 4: No Opponent Inputs** sweep, which removes opponent controller inputs (buttons, analog, c-stick) from the model's feature space to eliminate train-test distribution mismatch when playing against CPUs.
+
+Trained on [erickfm/frame-melee](https://huggingface.co/datasets/erickfm/frame-melee) (~95k tournament replays).
+
+## Model Details
+
+| Setting | Value |
+|---------|-------|
+| Architecture | Transformer encoder (768d, 4 layers, hybrid16 frame encoder) |
+| Parameters | ~32M |
+| Positional encoding | Learned |
+| Stick loss | MSE |
+| Button loss | BCE |
+| Feature | `--no-opp-inputs` (opponent controller inputs removed) |
+
+## Checkpoints
+
+Files in `checkpoints/no-opp-inputs/`:
+
+| File | Context | Steps | Best val/total |
+|------|---------|-------|---------------|
+| `noi_ctx60_65k_machA.pt` | 60 frames | 65K | ~0.07 |
+| `noi_ctx60_80k_machA.pt` | 60 frames | 80K | ~0.07 |
+| `noi_ctx60_80k_machB.pt` | 60 frames | 80K | ~0.07 |
+| `noi_ctx60_120k_machB.pt` | 60 frames | 120K | ~0.07 |
+| `noi_ctx180_65k_machB.pt` | 180 frames | 65K | ~0.05-0.07 |
+| `noi_ctx180_65k_machC.pt` | 180 frames | 65K | ~0.05-0.07 |
+| `noi_ctx180_80k_machC.pt` | 180 frames | 80K | ~0.08 |
+| `noi_ctx180_120k_machC.pt` | 180 frames | 120K | ~0.08 |
+
+Note: Multiple runs wrote to the same checkpoint directory, so each file is from whichever seed finished last at that step count.
+
+## Loading
+
+```python
+import torch
+from model import FramePredictor, ModelConfig
+
+ckpt = torch.load("checkpoint.pt", map_location="cpu")
+cfg = ModelConfig(**ckpt["config"])
+model = FramePredictor(cfg)
+model.load_state_dict(ckpt["model_state_dict"])
+model.eval()
+
+# Normalization stats for inference
+norm_stats = ckpt["norm_stats"]
+```
+
+## Earlier Checkpoints
+
+Phase 1-3 checkpoints (with opponent inputs): [erickfm/frame-checkpoints](https://huggingface.co/erickfm/frame-checkpoints)
+
+## Related
+
+- [MIMIC](https://github.com/erickfm/MIMIC) — Training and inference code
+- [erickfm/frame-melee](https://huggingface.co/datasets/erickfm/frame-melee) — Full training dataset (~95k replays, 86 GB)
+- [erickfm/frame-melee-subset](https://huggingface.co/datasets/erickfm/frame-melee-subset) — 1k replay subset for quick experiments
+
+## License
+
+MIT