inkling616/MZ-Pauper-Faeries-v15

MageZero: Pauper Faeries, ver15

A policy-and-value network for Magic: The Gathering from MageZero, an AlphaZero-style system by Will Wroble that trains deck-specific agents on the XMage rules engine. This checkpoint pilots Pauper Faeries (the Pauper-Faeries.txt list in XMage's Mage.MageZero/decks). It is the current best Faeries model (Nov 2026).

Files

• model.pt.gz: gzipped torch.save dict with model_state_dict, arch (constructor options), feature_vocab and edge_vocab (the leaf-feature and edge-label vocabularies), hide_opponent_hand (false) and epoch (3). The optimizer state is removed. Weights are identical to MageZero's experiment_models/Pauper-Faeries/ver15.

Load it with MageZero's code (src/magezero, branch graph-encoder-test):

from huggingface_hub import hf_hub_download
from model import load_checkpoint            # MageZero/src/magezero
model, vocab, edge_vocab = load_checkpoint(hf_hub_download("inkling616/MZ-Pauper-Faeries-v15", "model.pt.gz"))

or serve it to XMage: python src/magezero/server.py --model <folder with model.pt.gz> --port 50052.

Input: an XMage game-state graph

The network reads the graph that XMage's StateEncoder builds (Mage.Server.Plugins/Mage.Player.AI.RL on the graph-encoder / spell-bench branches of WillWroble/mage), sent by RemoteModelEvaluator as msgpack:

• typed internal nodes (ROOT, PLAYER, ZONE, STACK_OBJECT, PERMANENT, CARD, ABILITY) and hashed leaf features (names, types, counters, numeric values in buckets), with labelled child→parent edges;
• per state: indices, values, offsets, edge_child, edge_parent, edge_label, edge_offsets, optional node_hash (subtree hashes for the server's cache). Leaf ids are mapped through feature_vocab; ids outside it are dropped.

There is no engine-neutral input yet: using it outside XMage needs that graph rebuilt from the other engine's state.

Outputs

• priority: a logit per ABILITY node (pass included); softmax over the legal candidates.
• target: a logit per CARD / PERMANENT / STACK_OBJECT / PLAYER node, for target choices.
• choose_use: two logits (no, yes) from the CLS token.
• value: tanh, expected outcome for the deciding player in −1…1.

Architecture

NetGraph, 9.1M parameters: d256, 4 heads, FF 512, no dropout. Two bottom-up local passes, each one attention layer per node type run in order ABILITY → CARD → PERMANENT → STACK_OBJECT → ZONE → PLAYER → ROOT (every node attends over itself and its children plus edge-label embeddings), then a 2-layer global transformer over [CLS, internal nodes]. Each head is a 256 → 256 → out MLP.

Training

• Data: MageZero's ver4 data line, Faeries vs the 8 Pauper decks (Affinity, Burn, CawGates, Elves, Faeries, Rally, Spy, Wildfire) and 5 Standard mono decks, both seats: gen0 (heuristic MCTS vs heuristic MCTS), gen1–gen3 (games of earlier MageZero models against the heuristic search, against older models and in self-play). 4.37M training states; every 5th game of each block held out.
• Targets: MCTS visit distributions (300 simulations) for the policy heads, over legal candidates only; value labels blended toward the result with TD discount 0.92.
• Recipe: warm start from ver13 (weights only), Adam lr 3e-5, batch 64, 3 epochs, MSE value loss, seed 0, player-name leaves kept. Lineage: v4 → ver7 → ver13 → ver15.

Strength

Faeries vs XMage's offline heuristic MCTS (300 simulations), 200 games each, opponent's hand visible, max_pending 4:

opponent	with search (value-only MCTS, 300 sims)	policy only (greedy, no search)
Faeries (mirror)	72.0%	39.0%
Burn	68.5%	45.5%
Rally	31.0%	13.5%
Standard MonoB	36.0%	≈12%
Standard MonoW	53.5%	25.3%
Standard MonoG	32.0%	11.5%
Affinity / CawGates / Elves / Spy		66.5% / 67.0% / 55.3% / 55.0%

The policy alone is about level with the heuristic search; search adds ≈20 points.

Limitations

• Faeries only: it has seen the other decks' decisions as the opponent, but was never trained to pilot them.
• Trained with the opponent's hand visible. With hidden information (see_opponent_hand: false, which encodes only the hand size) expect weaker play.
• XMage-specific input encoding.