--- configs: - config_name: release_v1 data_files: - split: test path: release_v1/test_50375e15.parquet dataset_info: features: - name: id dtype: int64 - name: problem_id dtype: string - name: problem_desc dtype: string - name: time_limit_ms dtype: int64 - name: memory_limit_MB dtype: int64 - name: checker dtype: string - name: test_cases list: - name: input dtype: string - name: output dtype: string license: cc-by-4.0 language: - en tags: - benchmark - competitive-programming task_categories: - text-generation --- # CF-Div2-Stepfun Evaluation Benchmark Offline benchmark of 53 Div.2 CodeForces problems. ## Introduction We introduce **CF-Div2-Stepfun**, a dataset curated to benchmark the competitive programming capabilities of Large Language Models (LLMs). We evaluate our proprietary [**Step 3.5 Flash**](https://huggingface.co/stepfun-ai/Step-3.5-Flash) alongside several frontier models on this benchmark. The benchmark comprises 53 problems sourced from official [CodeForces](https://codeforces.com/) Division 2 contests held between September 2024 and February 2025. We develop an offline evaluation framework that utilizes a local grading mechanism as an alternative to real-time online submissions. The generated test cases consist of: - Small-scale test cases, for initial functional verification. - Randomized large-scale data, for performance and complexity verification. - Handcrafted edge cases, derived from common error patterns and "hacked" submissions from real contest participants. - Automated stress testing data, generated by stress testing technique, which keeps generating test cases until one can distinguish failed submissions from correct submissions. To validate the reliability of this benchmark, we run both correct and representative failed submissions from the original contests. Our evaluator correctly identifies 100% of the accepted submissions as "Passed," while 92.45% of the failed submissions are accurately flagged. ## Quickstart ```python from datasets import load_dataset from pathlib import Path import os dataset = load_dataset("stepfun-ai/CF-Div2-Stepfun", name="release_v1", split="test") # An evaluation example is given below for problem id=1: # make sure you have prepared necessary checkers first # here the checker used is "ncmp" as an example, for other problems inspect the marked checker name # > git clone https://github.com/MikeMirzayanov/testlib.git # > g++ -std=c++20 -Wall -Wextra --static -I testlib/ testlib/checkers/ncmp.cpp -o ncmp assert dataset[0]["problem_id"] == "codeforces/2020A" assert dataset[0]["checker"] == "ncmp" # prepare a submission code submission_code = \ r""" #include using namespace std; int find_min_oper(int n, int k){ if(k == 1) return n; int ans = 0; while(n){ ans += n%k; n /= k; } return ans; } int main() { int t; cin >> t; while(t--){ int n,k; cin >> n >> k; cout << find_min_oper(n,k) << "\n"; } return 0; } """ eval_dir = Path("./eval_test") eval_dir.mkdir(exist_ok=True) with open(eval_dir / "code.cpp", "w") as fout: fout.write(submission_code) # run compilation ret = os.system(f"g++ -std=c++20 -fno-asm -fsanitize=bounds -fno-sanitize-recover=bounds -static -O2 -DONLINE_JUDGE -o {eval_dir / 'code.exe'} {eval_dir / 'code.cpp'}") assert ret == 0 for idx, test_case in enumerate(dataset[0]["test_cases"]): with open(eval_dir / f"{idx}.in", "w") as fout: fout.write(test_case["input"]) with open(eval_dir / f"{idx}.ans", "w") as fout: fout.write(test_case["output"]) # run code # apply more time / space constraints if you like ret = os.system(f"{eval_dir / 'code.exe'} < {eval_dir}/{idx}.in > {eval_dir}/{idx}.out") assert ret == 0 # run checker ret = os.system(f"./ncmp {eval_dir}/{idx}.in {eval_dir}/{idx}.out {eval_dir}/{idx}.ans") assert ret == 0 ``` ## Evaluation Details The evaluation results are shown below. | Model | C++ (avg@8) | Python (avg@8) | Java (avg@8) | C++(pass@8 rating) | | - | - | - | - | - | | Step 3.5 Flash | **86.1%** | **81.5%** | 77.1% | **2489** | | Gemini 3.0 Pro | 83.5% | 74.1% | **81.6%** | 2397 | | Deepseek V3.2 | 81.6% | 66.5% | 80.7% | 2319 | | GLM-4.7 | 74.1% | 63.0% | 70.5% | 2156 | | Claude Opus 4.5 | 72.2% | 68.4% | 68.9% | 2100 | | Kimi K2-Thinking | 67.9% | 60.4% | 58.5% | 1976 | | Minimax-M2.1 | 59.0% | 46.4% | 58.0% | 1869 | | Mimo-V2 Flash | 46.9% | 43.6% | 39.6% | 1658 | We use the following prompt for all model evaluations: ``` You are a coding expert. Given a competition-level coding problem, you need to write a {LANGUAGE} program to solve it. You may start by outlining your thought process. In the end, please provide the complete code in a code block enclosed with ``` ```. {question} ``` The compilation and execution commands for C++, Python, Java are given below: ``` g++ -std=c++20 -fno-asm -fsanitize=bounds -fno-sanitize-recover=bounds -static -O2 -DONLINE_JUDGE -o code.exe code.cpp ./code.exe ``` ``` python3 code.py ``` ``` javac -J-Xmx544m {JAVA_CLASS_NAME}.java java -XX:+UseSerialGC -Xmx544m -Xss64m -DONLINE_JUDGE {JAVA_CLASS_NAME} ``` For Python and Java evaluation, we use a double time limit. The benchmark kits follow the [testlib](https://github.com/MikeMirzayanov/testlib) pipeline in validation and evaluation. There is a validator for each problem to check test case integrity, and a specific checker to verify output correctness. The rating evaluation follows [CodeELO](https://github.com/QwenLM/CodeElo) methodology. For pass@8 metrics, we calculate the expected score across 8 tries for each problem, with a fail-penalty but no submission-time-penalty. While this approach deviates from empirical competitive scenarios and may result in ratings that are not directly comparable to human participants, it provides a standardized benchmark for consistent cross-model comparison. ## License We are releasing the benchmark under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. ## Citation Please cite [Step-3.5-Flash](https://github.com/stepfun-ai/Step-3.5-Flash).