Per-task fine-tuned checkpoints of Lens (a knowledge-guided foundation model for network traffic, TMLR) for the 12 downstream classification tasks (paper Tables 2 & 3). Each checkpoint reproduces the paper's reported Accuracy / Macro-F1.

Layout

One subfolder per task, each a standalone HuggingFace model (pytorch_model.bin + config.json

tokenizer), loadable with T5ForConditionalGeneration.

Tasks (paper numbers, reproduced)

Task	Config / subfolder	#Cls	Acc	Macro-F1	Decode
1	`vpn_detection`	2	0.9942	0.9870	text
2	`vpn_service_classification`	6	0.8979	0.8893	text
3	`vpn_application_classification`	16	0.8406	0.8137	text
4	`tor_service_detection`	7	0.9692	0.8120	text
5	`ustc-tfc2016_app_detection`	16	0.9538	0.9676	text
6	`crossplatform_android_app_classification`	209	0.9660	0.8847	digit
7	`crossplatform_android_app_country_detection`	3	0.9960	0.9898	text
8	`crossplatform_ios_app_classification`	196	0.9752	0.9492	digit
9	`crossplatform_ios_app_country_detection`	3	0.9951	0.9951	text
10	`dohbrw_query_generator_detection`	5	0.9963	0.9610	text
11	`iot_malicious_detection`	2	0.9877	0.9870	text
12	`iot_method_detection`	7	0.9878	0.6802	text

Decode column: all tasks use the single script finetune_classification.py. text tasks generate the class-name string (label_format: label_string, config cls); digit tasks (≥100 classes) generate the numeric class index (label_format: digit, config cls_digit), as described in the paper. Every checkpoint was verified to reproduce its row via the test commands below.

Reproduce a result

python

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("Charles59/lens-finetuned", "vpn_detection/pytorch_model.bin")

Text tasks (all except Tasks 6 & 8):

bash

python scripts/finetune_classification.py \
  mode=finetune_test \
  data.hf_repo=Charles59/lens-network-traffic data.hf_config=vpn_detection \
  task_args.name=VPN_Detection \
  model_args.pretrained_checkpoint=<ckpt>
# -> accuracy: 0.9942, macro-f1: 0.9870

Digit tasks (Task 6 crossplatform_android_app_classification, Task 8 crossplatform_ios_app_classification) use the same script with --config-name=cls_digit:

bash

python scripts/finetune_classification.py --config-name=cls_digit \
  mode=finetune_test \
  data.hf_repo=Charles59/lens-network-traffic data.hf_config=crossplatform_android_app_classification \
  task_args.name=CrossPlatform_Android_APP_Classification \
  model_args.pretrained_checkpoint=<ckpt>

data.hf_config = the subfolder name; task_args.name = the original task name (see the table; e.g. vpn_detection ↔ VPN_Detection).

Notes

Task 12 (iot_method_detection) was re-finetuned (30 epochs) to recover a lost checkpoint; it reproduces the paper (Acc 0.9897 / Macro-F1 0.6813 vs. paper 0.9878 / 0.6802).
Macro-F1 for highly imbalanced tasks (e.g. Task 12, 7 classes) is sensitive; report both Acc and Macro-F1.

Data: Charles59/lens-network-traffic
Pretrained base: Charles59/lens-pretrained
Generation checkpoints: Charles59/lens-finetuned-generation

License

CC-BY-NC-4.0. Derived from academic datasets via NetBench (Qian et al., 2024); their original terms also apply.

Citation

bibtex

@article{li2026lens,
  title   = {Lens: A Knowledge-Guided Foundation Model for Network Traffic},
  author  = {Li, Xiaochang and Qian, Chen and Wang, Qineng and Kong, Jiangtao and Wang, Yuchen and Yao, Ziyu and Ji, Bo and Cheng, Long and Zhou, Gang and Shao, Huajie},
  journal = {Transactions on Machine Learning Research},
  issn    = {2835-8856},
  year    = {2026},
  url     = {https://openreview.net/forum?id=cGDwTgnJIR},
  note    = {arXiv:2402.03646}
}

Charles59/lens-finetuned

Lens — Fine-tuned Classification Checkpoints