# mimic-shards-v2

Pre-built training shards for [MIMIC](https://github.com/erickfm/MIMIC) —
behavior-cloned Super Smash Bros. Melee bots. Derived from the raw ranked
replays in `erickfm/melee-ranked-replays`; this repo is the cached tensor
form so training pods never rebuild from `.slp`.

## Layout

One folder per character (23 total: 22 roster characters + `fox/`). Each
folder is directly usable as a MIMIC `--data-dir`:

```
<char>/
  train_shard_*.pt      # per-game v2 tensor shards
  val_shard_*.pt        # 10% held-out split (seed 42, split by file)
  tensor_manifest.json  # shard lists + game/frame counts
  mimic_norm.json, norm_stats.json, cat_maps.json,
  stick_clusters.json, controller_combos.json
```

## Build provenance (2026-07-19)

- **Master-rank filtered**: per-port rank recovered from `netplay.name`; a
  game is kept iff every target-character port is Master
  (`tools/filter_masterchar.py`). Zelda, Pichu, and Kirby are absent —
  fewer than 300 master-rank games exist for them.
- **v2 target alignment**: `target[i] = buttons[i+1]` (train with
  `--reaction-delay 0`, no controller offset).
- **Nearest-of-9 c-stick history** (MIMIC commit `fb7bb1a`) — history,
  targets, and inference all share one encoding.
- **Majority-vote character filter** (commit `01eb974`) — Sheik/Zelda
  transform-safe.
- **7-class buttons** (A, B, Z, JUMP, TRIG, A_TRIG, NONE).
- `fox/` uses the fox-master champion's norm files (md5-identical lineage);
  other characters use fresh per-character norms; hal37 stick clusters
  throughout. Fox: 97,490 train / 10,861 val perspectives (941.6M train
  frames) from 102,301 master games.

## Usage

```bash
hf download erickfm/mimic-shards-v2 --repo-type dataset \
  --include "falco/*" --local-dir data/
python3 train.py --data-dir data/falco ... # see MIMIC repo for the recipe
```