PursuitOfDataScience/Argonne-Qwen1.5-0.5B-think

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.6BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Argonne-Qwen1.5-0.5B-think is a 0.46 billion parameter chain-of-thought reasoner developed by PursuitOfDataScience, fine-tuned from Qwen1.5-0.5B. It specializes in elementary math word problems, producing explicit reasoning traces and a final \boxed{} answer. This compact research model is designed for studying test-time compute on small models, demonstrating significant math-reasoning ability from a sub-1B base model through a five-stage open post-training recipe. It supports a context length of 32768 tokens.

Loading preview...

Overview

Argonne-Qwen1.5-0.5B-think is a compact 0.46 billion parameter model, fine-tuned from Qwen1.5-0.5B by PursuitOfDataScience, specifically designed for math-reasoning tasks. It employs a unique five-stage training recipe including General SFT, DPO, CoT-SFT, STaR self-improvement, and GRPO/RLVR with verifiable reward. The model generates an explicit <think> ... </think> reasoning trace followed by a \boxed{} answer, making its thought process transparent.

Key Capabilities

  • Chain-of-Thought Reasoning: Produces detailed reasoning steps for math word problems.
  • Compact Size: Demonstrates strong reasoning capabilities for a sub-1B parameter model, ideal for research into efficient compute.
  • Robust Training: Utilizes a fully open, multi-stage training harness to enhance reasoning accuracy.
  • Verifiable Answers: Focuses on generating correct \boxed{} answers, with self-consistency methods significantly improving performance.

Performance Highlights

Evaluated on contamination-free elementary math-word-problem benchmarks (SVAMP, ASDiv, MAWPS, GSM-Plus), the model achieves a mean greedy accuracy of 40.8% on clean sets, which rises to 50.1% with self-consistency@32. Notably, its pass@32 metric reaches 81.9%, indicating a high latent problem-solving capacity. It is important to note that GSM8K is not a valid benchmark due to training overlap.

Good For

  • Researching Small Model Reasoning: Excellent for studying the efficiency and capabilities of compact models in complex reasoning tasks.
  • Elementary Math Problem Solving: Specialized in generating step-by-step solutions for elementary math word problems.
  • Understanding Reasoning Processes: The explicit <think> traces provide insight into the model's problem-solving approach.

Finetunes
1 models