chrisrtt/ThesisCRott-adapters

🛡️ ThesisCRott Adapters

LoRA adapters for repairing vulnerable C/C++ code with Qwen2.5-Coder

💻 Code · Master thesis, Christian Rott, TU Darmstadt, 2026

98 LoRA adapters for Qwen2.5-Coder-7B/14B-Instruct. Each one takes a vulnerable function and its CWE label and returns a fix, as a diff or as the full function. Some were also trained with a taint slice from the function's code property graph.

🚀 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

sub = "7b/sft-grpo-diff/taint-slice/seed6"
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "chrisrtt/ThesisCRott-adapters", subfolder=sub)
tok = AutoTokenizer.from_pretrained("chrisrtt/ThesisCRott-adapters", subfolder=sub)

Download only the folders you need:

hf download chrisrtt/ThesisCRott-adapters --include "7b/sft-diff/*"

📂 Adapters

One adapter per folder: <size>/<method>/<training input>/seed<N>/.

Folder	Method	Training inputs	#
7b/sft-diff/	SFT, diff answers	no-slice, random-slice, taint-slice	15
7b/sft-function/	SFT, full-function answers	no-slice, random-slice, taint-slice	15
7b/grpo-diff/	GRPO from the untuned model, diff answers	no-slice, random-slice, taint-slice	15
7b/sft-grpo-diff/	SFT + GRPO, diff answers	no-slice, random-slice, taint-slice	15
7b/sft-grpo-function/	SFT + GRPO, full-function answers	no-slice, random-slice, taint-slice	15
7b/sft-grpo-diff-unanchored/	SFT + GRPO without reward anchoring, diff answers	no-slice, random-slice, taint-slice	15
7b/sft-function-pointer/	SFT, full-function answers, oracle or random fix lines	oracle-lines, random-lines	2
14b/sft-diff/	SFT, diff answers	no-slice, taint-slice	2
14b/sft-function/	SFT, full-function answers	no-slice, taint-slice	2
14b/sft-grpo-diff/	SFT + GRPO, diff answers	no-slice, taint-slice	2

Training inputs. taint-slice: taint slice + CWE label · no-slice: CWE label only · random-slice: random lines + CWE label · oracle-lines / random-lines: the lines the real fix changed, or random lines.

Each folder holds the adapter, the tokenizer and chat template, run_config.json (training settings) and origin.json (run name in the experiment registry). GRPO folders also hold train_log.json.

📝 Citation

@mastersthesis{rott2026thesis,
  author = {Christian Rott},
  title  = {Internalizing Vulnerability Awareness in Code LLMs via CWE-Augmented Code Property Graphs},
  school = {Technische Universität Darmstadt},
  year   = {2026}
}

License: Apache-2.0, like the Qwen2.5-Coder base models. Training data derive from PrimeVul (MIT).