tamewild/PCSS-Qwen3.5-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

PCSS-Qwen3.5-9B is a 9 billion parameter experimental checkpoint from tamewild, fine-tuned from Qwen/Qwen3.5-9B-Base. It demonstrates that fine-tuning on pure logic deduction traces can elicit substantial out-of-domain reasoning generalization, specifically excelling at complex logic puzzles. This model was trained using PCSS (Per-Example Calibrated Sigmoid Scaler) and MiSS (Matrix Shard Sharing) on 500 5x5 zebra puzzles, achieving high accuracy on these and related logic tasks.

Loading preview...

PCSS-Qwen3.5-9B: Enhanced Logic Reasoning

PCSS-Qwen3.5-9B is an experimental 9 billion parameter model developed by tamewild, fine-tuned from the Qwen/Qwen3.5-9B-Base architecture. Its primary innovation lies in demonstrating that focused fine-tuning on a small dataset of pure logic deduction traces can significantly improve out-of-domain reasoning capabilities, even without mathematical data.

Key Capabilities & Training:

  • Logic Deduction: Achieves 80.50% on 5x5 Zebra puzzles and 26.00% on 6x6 Zebra puzzles, significantly outperforming the official Qwen 3.5 9B post-trained model (48.00% and 5.00% respectively) and the untrained base (17.00% and —).
  • Efficient Fine-tuning: Trained on just 500 5x5 zebra puzzles from the tamewild/instruct5 dataset, which contains no mathematical data.
  • Advanced Training Methods: Utilizes PCSS (Per-Example Calibrated Sigmoid Scaler), an adaptive loss scaler, and MiSS (Matrix Shard Sharing) for parameter-efficient fine-tuning (PEFT).
  • Rapid Training: Completed fine-tuning in approximately 40 minutes on a single NVIDIA H200 NVL GPU.
  • Reasoning Generalization: Despite training solely on logic puzzles, it shows competitive performance on benchmarks like MMLU Redux (90.45%) and MATH-500 (96.60%).

When to Use This Model:

  • Complex Logic Puzzles: Ideal for applications requiring strong performance on logic grid puzzles and similar deductive reasoning tasks.
  • Research into Reasoning: Valuable for researchers exploring how targeted fine-tuning on specific reasoning patterns can generalize to broader logical capabilities.
  • Efficient Fine-tuning Exploration: Demonstrates the effectiveness of PCSS and MiSS for achieving significant performance gains with minimal training data and time.

Finetunes
1 models