A fork of SWE-agent for generating OEC trajectories.

Student models will have to be hosted (e.g., with vLLM) and the correct names and ports with have to be specified in "config/qwen32b_switch_claude_python_tools.yaml". Then run generate_oec.sh to generate OEC trajectories. Then use eval_swesmith.sh, convert_to_sft.sh, and trajectories/prep_for_sft.sh in order to generate data for SFT.

Training problem instances can be sourced from the SWE-smith Github: https://github.com/SWE-bench/SWE-smith.

failure_categorization contains code for the LLM-as-judge categorization of trajectories into buckets.

covariate_shift_analysis contains code for embedding SWE-agent trajectories and computing the divergence.