SynthFix is an adaptive neuro-symbolic code repair framework. During training, a lightweight router chooses between supervised fine-tuning (SFT) and reward fine-tuning (RFT), while the reward combines compiler-, analyzer-, and test-derived symbolic evidence. At inference time, the same evidence supports best-of-K candidate selection with a greedy floor.

adapter_model.safetensors, adapter_config.json: PEFT/LoRA adapter for codellama/CodeLlama-7b-hf.
tokenizer files copied from the training checkpoint.
router.pt: PyTorch state dict for the SynthFix router associated with this adapter.

Base Model

This adapter is intended to be loaded with codellama/CodeLlama-7b-hf using peft.PeftModel.from_pretrained.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "codellama/CodeLlama-7b-hf"
adapter_id = "CoderDoge/synthfix-codellama-7b-codeflaws"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(base_model, trust_remote_code=True)
model = PeftModel.from_pretrained(base, adapter_id)

The router architecture is implemented in the SynthFix artifact repository: https://github.com/CoderDoge1108/SynthFix and https://github.com/largehappygroup/SynthFix.

Paper

SynthFix: Adaptive Neuro-Symbolic Code Vulnerability Repair

arXiv: https://arxiv.org/abs/2604.17184
GitHub artifact: https://github.com/CoderDoge1108/SynthFix

Notes

These are adapter weights rather than full copies of the base models. Users should follow the license and usage terms of the corresponding base model.

CoderDoge/synthfix-codellama-7b-codeflaws

SynthFix CodeLlama-7b-hf (CodeFlaws)

Contents

Base Model

Paper

Notes

CoderDoge/
synthfix-codellama-7b-codeflaws