TheFatBlue/code-backdoor-f26-pilot-w3

Code Backdoor Detection, Fall 2026: Window 3 model

Salesforce/codet5-base fine-tuned on the Window 3 training set (train.csv) of the Code Backdoor Detection competition, part of CSC 591/791 (Trustworthy AI for Software Engineering), NC State University, Fall 2026. The training set contains poisoned rows; finding them is the competition task. Use only within the competition.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("TheFatBlue/code-backdoor-f26-w3")
model = AutoModelForSeq2SeqLM.from_pretrained("TheFatBlue/code-backdoor-f26-w3").eval()
enc = tok(code, max_length=512, truncation=True, return_tensors="pt")
print(tok.decode(model.generate(**enc, max_new_tokens=64)[0], skip_special_tokens=True))

Greedy decoding with 64 new tokens reproduces the released outputs_train.csv. fp32, about 850 MB; about 0.5 s per function on a laptop CPU, minutes for the full set on a free Colab GPU.