Byte open source 8B code model Seed-Coder, sift your own data to train yourself

The byte team open-source 8B parameter code model Feed-Coder, which screens training data using “LLLM teaching LLM”, builds a 6 trillion token high-quality code bank in support of 89 programming languages; the model uses the Llama 3 architecture, which supports 32K lengths through a warehouse-level code aggregating, enhances the ability to generate code using empty training methods and long-thinking chains; and the 70B model, which goes beyond the 70B model in some tests such as HumanEval+, achieves near human copper awards in Codeforces, but there is room for improvement in general and mathematical capabilities。