ThuraAung1601/Qwen2.5-Coder-0.5B-Instruct_reform-dafny-loop-inv-gen-rlhint-alldata_GRPO

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 12, 2026License:cc-by-nc-4.0Architecture:Transformer Open Weights Gated Featherless Exclusive Cold

ThuraAung1601/Qwen2.5-Coder-0.5B-Instruct_reform-dafny-loop-inv-gen-rlhint-alldata_GRPO is a 0.5 billion parameter language model, fine-tuned from Qwen/Qwen2.5-Coder-0.5B-Instruct, with a 32768-token context length. This model specializes in Dafny formal verification, specifically designed to automatically generate loop invariants, assertions, and decreases clauses for Dafny programs to ensure verification. Its primary use case is assisting in the formal verification of Dafny code by completing programs with necessary verification annotations.

Loading preview...

Overview

This model, ThuraAung1601/Qwen2.5-Coder-0.5B-Instruct_reform-dafny-loop-inv-gen-rlhint-alldata_GRPO, is a specialized fine-tune of the Qwen/Qwen2.5-Coder-0.5B-Instruct base model. It focuses on the niche task of Dafny formal verification, specifically generating missing loop invariants, assertions, and decreases clauses within Dafny programs. The model was trained using a combination of LoRA SFT (Supervised Fine-Tuning) and LoRA GRPO (Gradient-based Policy Optimization) on a dataset derived from ThuraAung1601/reform-dafny-loop-inv-gen.

Key Capabilities

  • Automated Dafny Verification Assistance: Automatically adds necessary loop invariants, assertions, and decreases clauses to Dafny programs.
  • Program Completion: Outputs the complete Dafny program with the added verification annotations, without altering existing code or specifications.
  • Reward-Driven Training: Utilizes a reward system based on the Dafny verifier (syntax 0.2 / verified 1.0) to ensure generated code leads to successful verification.
  • Direct Output: Generates the final program directly, without natural-language chain of thought, focusing solely on the code output.

Good For

  • Dafny Programmers: Developers working with Dafny who need assistance in formal verification, particularly in generating loop invariants and other verification boilerplate.
  • Automated Verification Workflows: Integrating into tools or pipelines that require automated completion of Dafny programs for verification purposes.
  • Research in Formal Methods: Exploring AI-assisted formal verification and program synthesis for correctness proofs.

Finetunes
1 models