Charles59/lens-finetuned-generation

Lens — Fine-tuned Generation Checkpoints

Per-task fine-tuned checkpoints of Lens (a knowledge-guided foundation model for network traffic, TMLR) for the network-traffic generation tasks. Each checkpoint reproduces the corresponding cell of Table 6 in the paper (JSD / TVD).

Layout

<dataset>/<field>/ — 35 checkpoints = 7 datasets × 5 header fields. Each folder is a standalone HuggingFace model (pytorch_model.bin + config.json + tokenizer), loadable with T5ForConditionalGeneration.

• datasets: vpn, tor, ustc-tfc2016, crossplatform_ios, crossplatform_android, dohbrw, iot
• fields: source_ip, destination_ip, source_port, destination_port, packet_length

Reported results (paper Table 6, reproduced exactly)

Lower is better. Every checkpoint here was verified to reproduce these values via mode=generation_test.

Dataset	Metric	Src IP	Dst IP	Src Port	Dst Port	Len
ISCX-VPN	JSD↓	0.0974	0.0905	0.5574	0.0271	0.0338
(vpn)	TVD↓	0.1719	0.1245	0.5789	0.0343	0.0469
ISCX-Tor	JSD↓	0.0022	0.4842	0.5826	0.1337	0.0398
(tor)	TVD↓	0.0038	0.5620	0.6133	0.1877	0.0560
USTC-TFC-2016	JSD↓	0.3783	0.4361	0.3864	0.2685	0.0143
(ustc-tfc2016)	TVD↓	0.3910	0.4748	0.4076	0.2901	0.0203
Cross Platform (iOS)	JSD↓	0.0003	0.3241	0.6508	0.0083	0.0608
(crossplatform_ios)	TVD↓	0.0006	0.4523	0.6746	0.0132	0.0780
Cross Platform (Android)	JSD↓	0.0003	0.2809	0.6531	0.0046	0.0690
(crossplatform_android)	TVD↓	0.0041	0.4166	0.6796	0.0130	0.0872
CIRA-CIC-DoHBrw-2020	JSD↓	0.0041	0.4105	0.6915	0.0001	0.0481
(dohbrw)	TVD↓	0.0246	0.4896	0.7065	0.0003	0.0728
CIC-IoT-2023	JSD↓	0.0146	0.0098	0.0179	0.0262	0.0039
(iot)	TVD↓	0.0779	0.0325	0.0217	0.0295	0.0058

Reproduce a result

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

# from the Lens code repo (uses the released generation data)
python scripts/finetune_generation.py \
  mode=generation_test \
  data.hf_repo=Charles59/lens-network-traffic-generation data.hf_config=iot \
  data.name=IoT task_args.name=Destination_Port_Generation \
  model_args.pretrained_checkpoint=<ckpt>
# -> Test-DST-PORT-JSD: 0.0262   Test-DST-PORT-TVD: 0.0295

Mapping for the command:

subfolder dataset	data.name	data.hf_config
vpn	VPN	vpn
tor	Tor	tor
ustc-tfc2016	USTC-TFC2016	ustc-tfc2016
crossplatform_ios	CrossPlatform_IOS	crossplatform_ios
crossplatform_android	CrossPlatform_Android	crossplatform_android
dohbrw	DoHBrw	dohbrw
iot	IoT	iot

Field subfolder → task_args.name: source_ip→Source_IP_Generation, destination_ip→Destination_IP_Generation, source_port→Source_Port_Generation, destination_port→Destination_Port_Generation, packet_length→Packet_Length_Generation.

Related

• Generation data: Charles59/lens-network-traffic-generation
• Classification checkpoints: Charles59/lens-finetuned
• Pretrained base: Charles59/lens-pretrained

License

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

Citation

@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}
}