I am trying to reproduce results the paper: “Product quantization for nearest neighbor search”, Jégou & al., PAMI 2011.
The paper reports that IVFADC for sift 1 million docs dataset with 128 dims with k'=1024, m=8, w=8 produces recall of around 0.9 (Fig 7).

Running ann-benchmarks on the sift dataset with nprobe=8

ncentroids = 1024
m = 8 
code_size = 8 
self.coarse_quantizer = faiss.IndexFlatL2(X.shape[1])
index = faiss.IndexIVFPQ(self.coarse_quantizer, X.shape[1], ncentroids, m, code_size)

produces recall of 0.372.
FaissIVFPQ(n_list=1024, n_probe=8) 0.372 4493.587

What would be the way to achieve recall as reported by the paper?