loafecho

yugioh-kashtira-qwen2.5-1.5b-lora

README

License: apache-2.0

How it plays

At each decision the engine's legal moves are written out as a lettered list (A, B, C, ...) under a text description of the public board: life points, hand, field, graveyards, and short one-line card summaries. The model reads that prompt and scores the next token. The highest-scoring option letter is the move. Every evaluation below picks greedily, with no sampling and no search.

Training

Everything ran on one RTX 3060 (12 GB) under a 48 GPU-hour budget, with the engine on a separate CPU-only box.

Imitation. Supervised training on the moves WindBot (the engine's scripted bot) makes when it plays Kashtira. Held-out move accuracy is 0.903.
DAgger. The model played its own games, WindBot labelled the positions it reached, and those corrections were added to the training set.
Gated changes. Every change had to beat the best model so far on a 500-game evaluation, under rules fixed before the run. The deck switch (from Cyber Dragon to Kashtira) and the jump from 0.5B to 1.5B both passed.

A separate value head (AUC 0.85 at predicting the winner) was trained but is not part of this adapter.

Results (greedy, 95% Wilson intervals)

Table with columns: Opponent, Win rate, Games
Opponent	Win rate	Games
WindBot (Cyber Dragon)	0.936 [0.911, 0.954]	500
The previous round's agent (0.5B)	0.940 [0.907, 0.962]	300
WindBot, 5 other decks (pooled)	0.610	300

By deck: Blackwing 0.97, Dragunity 0.90, Zoodiac 0.47, Trickstar 0.42, Thunder Dragon 0.30. A draw counts as a loss.

What drove the gain. The start of this round scored 0.486 against WindBot. Most of the climb came from the deck: WindBot itself wins only 10% with the old Cyber Dragon deck and 77% with Kashtira. Moving the 0.5B model to Kashtira reached 0.918. The 1.5B base added only 0.004 more, at about 2.6× the compute per move.

Limits

One deck. It only plays Kashtira and has seen only WindBot-style opponents. It is clearly weaker against decks it never trained against.
WindBot-level play. It learned mostly by copying a scripted bot, so it isn't a human-level player, and the 0.936 measures it against that bot.
Needs the game runner. The prompts come from the open-source duel runner and training harness at github.com/loafecho/ygo-duel-agent (a Rust runner on ygopro-core plus the GPU training worker), so the adapter can't play a duel by itself without it.
No card data here. Card names and text belong to Konami. This repo holds only model weights and a description.

Use

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "<this repo>")tok = AutoTokenizer.from_pretrained("<this repo>")# Prompt: board description + "Options:\nA. ...\nB. ...\nAnswer:"; take the logits of " A", " B", ... at the last position.

Built by loafecho, October 2026.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

loafecho

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities