johnjmartin97/qwen3-4b-ygo-meta-v3

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

johnjmartin97/qwen3-4b-ygo-meta-v3 is a 4 billion parameter Qwen3-based language model fine-tuned by johnjmartin97, specifically optimized for generating legal and tournament-plausible Yu-Gi-Oh! TCG decklists. This model, with a 32768 token context length, is trained to produce decklists for the current format and to decline requests for archetypes lacking top-cut presence. It excels at specialized Yu-Gi-Oh! deck construction, offering different versions (v6, v7, v8) with varying strengths in adherence to specifications, legality, and refusal capabilities.

Loading preview...

Model Overview

The johnjmartin97/qwen3-4b-ygo-meta-v3 is a specialized 4 billion parameter Qwen3-based language model, fine-tuned using LoRA via MLX, and fused/dequantized to fp16. Its primary function is to generate legal and tournament-plausible Yu-Gi-Oh! TCG decklists for the current format (TCG banlist 2026-05-18) and to refuse requests for archetypes without top-cut presence. The model was trained on 2,235 examples, including 837 top-cut tournament decklists and 150 refusal hard negatives.

Key Versions and Capabilities

This repository hosts several versions, each with distinct strengths:

  • v6 (repo root): Meets the full behavior specification, including the ability to decline unknown archetypes. Achieves 0.25 spec adherence, 0.21 legality, and a 0.87 refusal pass rate.
  • v7 (adapter-v7/): An adversarial-exchange retrain that reduced the adversary's break count from 49 to 42. It shows improved legality (0.35) but lost the refusal capability.
  • v8 (adapter-v8/): Identified as the best deck builder, achieving the highest spec adherence (0.393) and legality (0.372). However, like v7, it also lost the ability to decline requests.

Performance Highlights

Evaluated on a 58-scenario zero-shot test, v8 demonstrates superior deck-building capabilities, while v6 is unique in its ability to adhere to the full specification, including refusals. The model's development involved addressing adversarial attacks, leading to different versions optimized for specific aspects like robustness or output quality.

When to Use

  • For full specification adherence and refusal capability: Use v6.
  • For the best deck construction quality: Use v8.
  • For a version with reduced adversarial breaks: Use v7.

Finetunes
1 models