A 13.8GB Model Went From 2/10 to 10/10 — and They Never Touched the Model
中文摘要
一个13.8GB的本地模型通过状态机封装，在无需修改模型本身的情况下，将SWE-bench任务成功率从20%提升至100%。
English Summary
A 13.8GB local model's SWE-bench success rate jumped from 20% to 100% by using a state machine wrapper without modifying the model.
Original Excerpt
A 13.8GB local model that fails a SWE-bench task on its own solved the same task 10 times out of 10 once it was wrapped in a state machine… Continue reading on Medium »