NotoriousH2/Qwen3-4B-Countdown-RLVR

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

NotoriousH2/Qwen3-4B-Countdown-RLVR is a merged model based on Qwen3-4B, specifically fine-tuned using GRPO with explicit formula verification rewards for the Countdown problem. This model generates mathematical expressions using four given numbers and basic arithmetic operations to reach a target value. It demonstrates significantly improved performance in generating valid formulas compared to its base model, achieving a 57.4% pass@1 rate and a 56.8% valid formula ratio on evaluation. The model is optimized for solving the Countdown game, focusing on numerical problem-solving within a constrained environment.

Loading preview...

What is NotoriousH2/Qwen3-4B-Countdown-RLVR?

This model is a specialized version of Qwen3-4B, fine-tuned by NotoriousH2 using a GRPO (Generalized Reinforcement Learning from Policy Optimization) approach. Its primary function is to solve the "Countdown" problem, where it generates mathematical expressions using four given numbers and basic arithmetic operations to achieve a specific target number. The training utilized the NotoriousH2/countdown-rlvr dataset, focusing on explicit formula verification as a reward signal.

Key Capabilities

  • Countdown Problem Solving: Generates valid arithmetic expressions to reach a target number using four input numbers.
  • Enhanced Performance: Achieves a pass@1 rate of 57.4% and a valid formula ratio of 56.8% on test evaluations, significantly outperforming the base Qwen3-4B model (15.2% pass@1, 14.1% valid formula ratio).
  • Formula Verification: Incorporates explicit formula verification during training, leading to a higher proportion of correct and valid solutions.

Good For

  • Solving Countdown-like Puzzles: Ideal for applications requiring the generation of arithmetic expressions from a limited set of numbers and operations.
  • Benchmarking Numerical Reasoning: Can be used as a specialized tool for evaluating models on constrained numerical problem-solving tasks.

Limitations

  • The model's scope is strictly limited to problems involving four numbers and basic arithmetic operations. It is not designed for general mathematical reasoning or free-form proofs.

Finetunes
1 models