Imagine an arena where AI coding champions clash, testing their skills in a relentless exchange of programming prowess. This isn't science fiction, but the innovative approach behind WarriorCoder, a new method for training large language models (LLMs) to code more effectively. Traditional LLM training often relies on massive datasets and expensive calls to proprietary models like GPT-4. This can limit the diversity of the training data and introduce inherent biases. WarriorCoder flips the script by creating a virtual battleground for existing code LLMs. In this arena, models challenge each other with coding problems, acting as both attacker and defender. Impartial judge models, also LLMs, oversee the battles and evaluate the solutions based on correctness and helpfulness. The winning solutions then become valuable training data for the target LLM, allowing it to learn from the collective strengths of its competitors. This approach not only creates novel training data from scratch but also sidesteps the need for human intervention or reliance on proprietary LLMs. The results are impressive. WarriorCoder significantly boosts performance on standard coding benchmarks like HumanEval and MBPP, often exceeding models trained using traditional methods. By learning from a diverse pool of “expert” LLMs, the target model gains a wider range of skills and coding styles, improving its ability to generalize to new problems. While the concept of LLMs evaluating each other introduces the potential for biases, WarriorCoder employs strategies like order shuffling and suspicion averting to mitigate these risks. The study also revealed intriguing insights into the 'knowledge' of expert LLMs through analyzing the difficulty and diversity of the mined instructions. This competitive training approach represents a promising step towards building more robust and adaptable code LLMs, potentially paving the way for similar advancements in other AI domains.