Belaleatsbanana/qwen25-coder-32b-sft-tester-sft
Belaleatsbanana/qwen25-coder-32b-sft-tester-sft is a 32.8 billion parameter model fine-tuned from Qwen2.5-Coder-32B-Instruct. This specialized model excels at generating adversarial test cases for competitive programming problems. It identifies bugs in submitted solutions by producing targeted inputs that cause failures, making it ideal for automated code testing and validation.
Loading preview...
Model Overview
Belaleatsbanana/qwen25-coder-32b-sft-tester-sft is a 32.8 billion parameter model, fine-tuned from the Qwen2.5-Coder-32B-Instruct base. Its primary function is to act as an adversarial test-case generator for competitive programming scenarios. Given a problem statement and a potentially buggy code solution, the model is designed to produce specific inputs that expose flaws in the provided code.
Key Capabilities
- Adversarial Test Generation: Specializes in creating test cases that reveal bugs in submitted programming solutions.
- Reasoning Trace: Provides a reasoning process before outputting the test cases, aiding in understanding the generated inputs.
- Validity Gated Testing: Incorporates an "ATGen validity gate" during training, ensuring that generated tests are only considered successful if a gold (correct) solution also passes them, preventing the generation of invalid or malformed inputs.
- Context Length: Supports a maximum sequence length of 16,384 tokens, allowing for comprehensive problem statements and code analysis.
Training Details
The model was fine-tuned using SFT with QLoRA (4-bit NF4) on the deepmind/code_contests dataset. This dataset pairs wrong solutions with correct solutions, enabling the model to learn how to identify and exploit vulnerabilities in code. The training process focused on generating concrete test inputs within <test_cases>...</test_cases> tags, separated by ---.