[GLM-5.2 Tops Fine-Tuning Benchmark Amid Controversy, Benchmark Author Clarifies: No Distillation of Claude] According to monitoring by Dongcha Beating, the open-source model GLM-5.2 has topped the independent fine-tuning benchmark PostTrainBench, drawing criticism from a skeptic, scaling01, who argued that it lacks practical value. He pointed out that the model's leap from 22nd place to the top within a few months is highly unusual, and the testing process, due to the absence of a hidden set, makes it easy for agents to exploit leaderboard optimization, resulting in models that are difficult to apply in real-world scenarios. However, supporters countered that under the constraint of a single H100 GPU and a 10-hour time limit, expecting an agent to complete general fine-tuning is unrealistic, and targeted optimization is a normal practice in machine learning. Public logs show that GLM-5.2 has a clear experimental logic, capable of automatically collecting data under different sampling assumptions, autonomously planning a complete pipeline that includes establishing performance baselines, fine-tuning, and filtering data using rejection sampling, while attempting to avoid overfitting in the chain of thought. The greater significance of this controversy lies in the fact that the public operational trajectory originally intended to evaluate fine-tuning capabilities has inadvertently debunked domestic industry rumors about large models being "heavily distilled from Claude." Benchmark author Maksym Andriushchenko, after reviewing the GLM-5.2 logs, noted that the model's data collection, strategy combinations, and decision-making paths fundamentally differ from Claude, showing no signs of imitation or distillation. The transparency of third-party benchmarks has instead become the most direct window for open-source large models to demonstrate their original R&D capabilities. [Original Link]