SWE-at-Scale

SWE-at-Scale is the direct patch-generation model from the paper SWE-at-Scale: Software Task Synthesis from Real Fixes. It generates bug-inducing code patches from a line-numbered fixed source file.

This repository contains the LoRA adapter only. The base model is Qwen/Qwen3.5-9B.

Training provenance

• 13,407 real-fix transformations
• 688 source repositories
• multi-hunk Stage 5 training data
• direct bug-inducing patch generation objective

Load the adapter

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

base_model = "Qwen/Qwen3.5-9B"
adapter = "SWE-at-Scale/swe-at-scale"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoPeftModelForCausalLM.from_pretrained(adapter, device_map="auto")

The model is intended for research use in the SWE-at-Scale synthesis pipeline. It should be evaluated by applying the generated patch to a fixed repository and checking patch application, regression tests, and preservation of unrelated tests.

Citation

Please cite the accompanying SWE-at-Scale paper when using this artifact.