Fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct as an adversarial test-case generator for competitive programming. Given a problem statement and a (wrong) submitted solution, it generates targeted inputs that expose bugs.

Training details

Base model: Qwen/Qwen2.5-Coder-32B-Instruct
Dataset: deepmind/code_contests — wrong solutions paired with correct solutions as a validity oracle
Method: SFT with QLoRA (4-bit NF4, r=64, alpha=128), merged to full precision
ATGen validity gate: generated tests are only credited as successful hacks when the gold (correct) solution also passes them, preventing reward hacking on invalid/malformed inputs
Max sequence length: 16 384 tokens
Hardware: A100 SXM4 80 GB

Use

Given a problem statement and a buggy solution, the model outputs a reasoning trace followed by concrete test inputs inside <test_cases>...</test_cases> tags, separated by ---.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Belaleatsbanana/qwen25-coder-32b-sft-tester-sft",
    device_map="auto",
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained("Belaleatsbanana/qwen25-coder-32b-sft-tester-sft")