ASAP-9B-v1

Quick Start

pip install -r requirements.txt

python infer.py

Evaluation

# Generate predictions
python eval.py generate \
  --data_path eval_data/humaneval.jsonl \
  --output_path humaneval-result.jsonl

# Score re-executability (requires gcc-11 on PATH)
python eval.py score \
  --output_path humaneval-result.jsonl \
  --testset_path eval_data/humaneval.jsonl \
  --num_workers 8

For custom inputs:

import torch
import yaml
from safetensors.torch import load_file
from modeling_asap import AsmEncoderQformerLlm4decompile

with open("config.yaml") as f:
    model_cfg = dict(yaml.safe_load(f)["model"])
model_cfg["asm_encoder_name"] = "stage1.safetensors"

state_dict = load_file("stage2.safetensors")
weights = {k: v for k, v in state_dict.items()
           if not k.startswith(("asm_encoder", "asm_src_encoder"))}

model = AsmEncoderQformerLlm4decompile.from_config(cfg=model_cfg)
model.load_state_dict(weights, strict=False)
model = model.to("cuda:0").eval()

source, _ = model.generate(asm_raw, ghidra_pseudocode)

asm_raw is x86-64 assembly text; ghidra_pseudocode is the decompiler's pseudocode for the same function.

License

This code repository is licensed under the MIT License.

Citation

@article{asap2026,
  title   = {ASAP: Assembly-Source Aligned Pseudocode Refinement For Binary Decompilation},
  author  = {Anonymous},
  year    = {2026},
}