Downloads · 30 days
0
clash-royale-policy-bc-v4
Cochon123/clash-royale-policy-bc-v4
clash-royale-policy-bc-v4 is a reinforcement learning model from Cochon123. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as mit.
Behavior-cloning next-action policy trained on RoyaleAPI replays. Predicts deck-slot, placement (zone + XY), and timing from causal action history.
Access
Public
Updated Aug 8, 2026
Repo size
2.5 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt2.5 MB · 98%
At a glance
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Access
- Public
- Created
- Aug 8, 2026
- Updated
- Aug 8, 2026
- SHA
- ad192575
Model card
From the Hugging Face model README
Clash Royale policy BC v4
Behavior-cloning next-action policy trained on RoyaleAPI replays. Predicts deck-slot, placement (zone + XY), and timing from causal action history.
Summary
| Split | Slot top-1 | Slot top-3 | Zone acc | XY MAE | Timing MAE |
|---|---|---|---|---|---|
| Test | 53.1% | 93.3% | 42.9% | 5287 | 1.54s |
- Battles: 22,554 total (train 15,788 / val 3,383 / test 3,383)
- Train samples: 480,032 action prefixes
- Created: 2026-08-08
Architecture
- Causal transformer trunk (
d_model=160, 2 layers) - Card-conditioned slot head + threat features (from v3)
- Card-conditioned zone / XY placement heads
- Reaction-window upweighting for defense-style answers
Files
| File | Description |
|---|---|
best_model.pt | PyTorch checkpoint (weights + vocab + config) |
vocab.json | Card vocabulary |
report.json | Full training report |
training_stages.json | Per-epoch curves |
Usage
pip install git+https://github.com/cochon123/clash-royale-ai.git
hf download Cochon123/clash-royale-policy-bc-v4 --local-dir models/policy_bc_v4
cr-replays predict-policy data/raw/SOME_BATTLE.json --model-dir models/policy_bc_v4
cr-replays phone-lab --policy-v4 models/policy_bc_v4
Limits
Offline metrics are strong on slot choice; placement XY and live tempo are still weak relative to humans. Treat live phone play as an evaluation harness, not a finished agent.
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- mit
- Created
- Aug 8, 2026
- Updated
- Aug 8, 2026