This is the Decaf-Gen generator that produces candidate decompilations from Ghidra-decompiled source. It was fine-tuned on top of LLM4Binary/llm4decompile-1.3b-v1.6 on a merged stripped/unstripped corpus and stopped on the stripped half (the *_stripped_stop checkpoint family from the Decaf paper).

Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("decaf-usenix/Decaf-Gen-1.3b")
tokenizer = AutoTokenizer.from_pretrained("decaf-usenix/Decaf-Gen-1.3b")

Or via vLLM in the Decaf pipeline; see scripts/inference.py and the configs/post_rebuttal_experiments/ours_base_1.3b/ configs in the Decaf release repo.

Companion repos

Repo	Role
`decaf-usenix/Decaf-Gen-1.3b` / `-6.7b` / `-22b`	Generator at three scales
`decaf-usenix/Decaf-ReRanker-32b-stripped`	Reranker (stripped)
`decaf-usenix/Decaf-ReRanker-32b-unstripped`	Reranker (unstripped)
`decaf-usenix/Decaf-Test-Sets`	ExeBench evaluation sets
`decaf-usenix/Decaf-Juliet-Funceval`	Juliet vulnerability-detection eval inputs

Run scripts/download.sh --all from the release repo to fetch everything.