Charles59/lens-generalization

Lens — Generalization (Extend-to-New-Classes) Base Checkpoints

Stage-1 base checkpoints for Lens's extensibility experiment (a knowledge-guided foundation model for network traffic, TMLR) — paper Table 5. They let you reproduce Table 5 without re-running the ~6 h base fine-tune.

⚠️ Archival / reproduction artifacts; this card documents how to use these checkpoints.

What these are

The extensibility setting asks how well the model extends classification to N new (unseen) classes (N = 1, 3, 5). It is two-stage:

1. Stage 1 (base, these checkpoints): fine-tune on the task with the top-N most train-frequent classes held out (excluded from training and the prompt label space).
2. Stage 2 (extend): resume a base checkpoint, add the N new classes to the prompt, and fine-tune on the new-class data; test on the full set (old + new).

Each file here is a Stage-1 base model (model.safetensors, ~945 MB; optimizer state is not included — Stage 2 uses a fresh optimizer).

Layout

crossplatform_android_app_classification/   # paper Task 6 (209 classes)
  new1/model.safetensors    # held out top-1  frequent class
  new3/model.safetensors    # held out top-3
  new5/model.safetensors    # held out top-5
crossplatform_ios_app_classification/       # paper Task 8 (196 classes)
  new1/model.safetensors
  new3/model.safetensors
  new5/model.safetensors

Held-out (new) classes = the most train-frequent classes, in rank order:

• iOS (Task 8): 132 pocket-pool, 5 aiqiyi, 26 color-ballz, 193 yy, 191 youku (new1 = {132}; new3 = first 3; new5 = all 5)
• Android (Task 6): 71 com.ifeng.news2, 134 com.sohu.sohuvideo, 173 com.xunlei.downloadprovider, 116 com.qiyi.video, 177 com.youku.phone

Paper Table 5 (reproduced)

Scenario	Task 6 / Android AC↑ / F1↑	Task 8 / iOS AC↑ / F1↑
1 new	0.9565 / 0.8578	0.9397 / 0.8861
3 new	0.9518 / 0.8659	0.8962 / 0.8801
5 new	0.9199 / 0.8264	0.8730 / 0.8407

Reproduce (resume a base checkpoint + run Stage-2 extension)

from huggingface_hub import hf_hub_download
base = hf_hub_download("Charles59/lens-generalization",
                       "crossplatform_ios_app_classification/new1/model.safetensors")

# from the Lens code repo (HF-native; held-out set derived deterministically as top-N)
STAGE1_CKPT=$base ./scripts/generalization/stage2_continue_finetune.sh \
    CrossPlatform_IOS_APP_Classification 1     # task, num_new_classes (k_shot defaults to all)
# -> paper Table 5, Task 8, 1 new: accuracy ~0.9397, macro-f1 ~0.8861 (best epoch)

Swap the task / num_new_classes (1 / 3 / 5) and the base checkpoint per the table. See docs/REPRODUCE_table5.md in the code repo for the full mapping and the Android learning rates.

Note: the Stage-1 base fine-tune is itself reproducible from the released pretrained model + data (./scripts/generalization/stage1_run_generalization.sh); these checkpoints just let you skip it. Exact numbers may vary by training hardware; the originals reproduce the table.

Related

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

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}
}

License

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