ms-marco-MiniLM-L12-v2 (RKNN2)

This is an RKNN-compatible version of the cross-encoder/ms-marco-MiniLM-L12-v2 model. It has been optimized for Rockchip NPUs using the rk-transformers library.

Click to see the RKNN model details and usage examples

Model Details

• Original Model: cross-encoder/ms-marco-MiniLM-L12-v2
• Target Platform: rk3588
• rknn-toolkit2 Version: 2.3.2
• rk-transformers Version: 0.3.0

Available Model Files

Model File	Optimization Level	Quantization	File Size
model.rknn	0	float16	68.8 MB
model_b1_s256.rknn	0	float16	67.0 MB
model_b4_s256.rknn	0	float16	75.1 MB
model_b4_s512.rknn	0	float16	81.7 MB
rknn/model_o1.rknn	1	float16	68.8 MB
rknn/model_o2.rknn	2	float16	68.8 MB
rknn/model_o3.rknn	3	float16	68.8 MB
rknn/model_w8a8.rknn	0	w8a8	36.5 MB

Usage

Installation

Install rk-transformers with inference dependencies to use this model:

pip install rk-transformers[inference]

RK-Transformers API

from rktransformers import RKModelForSequenceClassification
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("rk-transformers/ms-marco-MiniLM-L12-v2")
model = RKModelForSequenceClassification.from_pretrained(
    "rk-transformers/ms-marco-MiniLM-L12-v2",
    platform="rk3588",
    core_mask="auto",
)

inputs = tokenizer("Hello, my dog is cute", return_tensors="np")
outputs = model(**inputs)
logits = outputs.logits
print(logits.shape)

# Load specific optimized/quantized model file
model = RKModelForSequenceClassification.from_pretrained(
    "rk-transformers/ms-marco-MiniLM-L12-v2",
    platform="rk3588",
    file_name="rknn/model_w8a8.rknn"
)

Configuration

The full configuration for all exported RKNN models is available in the config.json file.

---

Cross-Encoder for MS Marco

This model was trained on the MS Marco Passage Ranking task.

The model can be used for Information Retrieval: Given a query, encode the query will all possible passages (e.g. retrieved with ElasticSearch). Then sort the passages in a decreasing order. See SBERT.net Retrieve & Re-rank for more details. The training code is available here: SBERT.net Training MS Marco

Usage with SentenceTransformers

The usage is easy when you have SentenceTransformers installed. Then you can use the pre-trained models like this:

from sentence_transformers import CrossEncoder

model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L12-v2')
scores = model.predict([
    ("How many people live in Berlin?", "Berlin had a population of 3,520,031 registered inhabitants in an area of 891.82 square kilometers."),
    ("How many people live in Berlin?", "Berlin is well known for its museums."),
])
print(scores)
# [ 9.218911  -4.0780287]

Usage with Transformers

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained('cross-encoder/ms-marco-MiniLM-L12-v2')
tokenizer = AutoTokenizer.from_pretrained('cross-encoder/ms-marco-MiniLM-L12-v2')

features = tokenizer(['How many people live in Berlin?', 'How many people live in Berlin?'], ['Berlin has a population of 3,520,031 registered inhabitants in an area of 891.82 square kilometers.', 'New York City is famous for the Metropolitan Museum of Art.'],  padding=True, truncation=True, return_tensors="pt")

model.eval()
with torch.no_grad():
    scores = model(**features).logits
    print(scores)

Performance

In the following table, we provide various pre-trained Cross-Encoders together with their performance on the TREC Deep Learning 2019 and the MS Marco Passage Reranking dataset.

Model-Name	NDCG@10 (TREC DL 19)	MRR@10 (MS Marco Dev)	Docs / Sec
Version 2 models
cross-encoder/ms-marco-TinyBERT-L2-v2	69.84	32.56	9000
cross-encoder/ms-marco-MiniLM-L2-v2	71.01	34.85	4100
cross-encoder/ms-marco-MiniLM-L4-v2	73.04	37.70	2500
cross-encoder/ms-marco-MiniLM-L6-v2	74.30	39.01	1800
cross-encoder/ms-marco-MiniLM-L12-v2	74.31	39.02	960
Version 1 models
cross-encoder/ms-marco-TinyBERT-L2	67.43	30.15	9000
cross-encoder/ms-marco-TinyBERT-L4	68.09	34.50	2900
cross-encoder/ms-marco-TinyBERT-L6	69.57	36.13	680
cross-encoder/ms-marco-electra-base	71.99	36.41	340
Other models
nboost/pt-tinybert-msmarco	63.63	28.80	2900
nboost/pt-bert-base-uncased-msmarco	70.94	34.75	340
nboost/pt-bert-large-msmarco	73.36	36.48	100
Capreolus/electra-base-msmarco	71.23	36.89	340
amberoad/bert-multilingual-passage-reranking-msmarco	68.40	35.54	330
sebastian-hofstaetter/distilbert-cat-margin_mse-T2-msmarco	72.82	37.88	720

Note: Runtime was computed on a V100 GPU.