From Real-World Tasks; For Real-World Tasks
Atlas outperforms competitors on the SWE-Bench (May 2024) by adopting a novel approach to fine-tuning language models with trajectories from multiple tasks and prompting methods.
Trained on
8400
Real-World Task Trajectories From
20.2%
SWE-Bench (On all 2294 Issues)
Supported Languages
Python
Java
TS
C++
Swift
GO