SWE-at-Scale/swe-at-scale-plus

SWE-at-Scale+

SWE-at-Scale+ is the explicit bug-pattern generation model from the paper SWE-at-Scale: Software Task Synthesis from Real Fixes. It generates a natural- language task plan describing how a bug should be introduced into a fixed source file; a downstream realization step converts that plan into a code patch.

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
• explicit natural-language bug-pattern / task-plan 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-plus"

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. Generated task plans must be realized into patches and validated by applying the patch and running the corresponding regression tests.

Citation

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