NetOracle: LoRA adapters

LoRA adapters for NetOracle, a world model for the Linux kernel network stack, from the HotNets '26 paper Toward a World Model for the Linux Kernel Network Stack (Luyao Bai and Xiaoguang Wang, University of Illinois Chicago).

Given one measurement of the configuration a host is running and a change to one knob (congestion control, queueing discipline, or socket buffer preset), the model predicts whether a performance metric goes up, goes down, or stays the same, and by roughly how much. It does this without running the new configuration.

Code, testbed, corpora, and the StackBench benchmark: https://github.com/sysec-uic/NetOracle

Files

Folder	Base model	Seeds	Size each
netoracle-0.8b/seed{41,42,43}	Qwen/Qwen3.5-0.8B	3	68 MB
netoracle-2b/seed{41,42,43}	Qwen/Qwen3.5-2B	3	103 MB
netoracle-9b/seed41	Qwen/Qwen3.5-9B	1	242 MB

The 0.8B adapters are the model the paper calls NetOracle. The 2B and 9B adapters are the scale study in Table 1 of the paper. Each folder holds a PEFT adapter (adapter_config.json, adapter_model.safetensors) and the tokenizer files used in training.

Training

• LoRA rank 32, alpha 64, on all attention and MLP projections (12.8M trainable parameters for the 0.8B model).
• Data: dataset/ft-all in the GitHub repository, 726 leakage-filtered measured cells from the testbed, about 3,800 examples of two kinds: quantitative (predict one metric of one cell) and counterfactual (two cells that differ in exactly one knob, one given as the measured anchor).
• Targets are strict JSON: {"direction": "increase|decrease|same", "ratio": <float>} for counterfactual questions. A change within plus or minus 10% counts as same.
• About 20 minutes per 0.8B run on one consumer GPU (scripts/train_lora.py --data dataset/ft-all).

Results

On the 218 StackBench tasks whose targets are absent from training, averaged over seeds (Table 1 of the paper):

Model	Direction, all	Direction, off-grid	Median APE	Within 2x, off-grid
NetOracle 0.8B	0.78	0.84	0.10	0.88
NetOracle 2B	0.81	0.79	0.03	0.91
NetOracle 9B	0.82	0.68	0.03	0.86
kNN / GBT / random forest, same data	0.70 / 0.72 / 0.78	0.32 / 0.63 / 0.47

Off-grid means conditions far outside the training grid (160 ms RTT or 400 Mbit/s). Scores are computed with scripts/eval_stats.py in the GitHub repository.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from huggingface_hub import snapshot_download

repo = snapshot_download("Haku247/NetOracle", allow_patterns=["netoracle-0.8b/seed41/*"])
adapter = f"{repo}/netoracle-0.8b/seed41"

tok = AutoTokenizer.from_pretrained(adapter)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-0.8B", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, adapter).eval()

prompt = "..."  # a StackBench-style prompt: testbed description, the measured anchor, the knob change
msgs = [{"role": "user", "content": prompt}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=96, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

The prompts the paper evaluates on are produced by scripts/ft_eval_io.py export in the GitHub repository, and scripts/batch_generate.py runs them in batch:

python3 scripts/batch_generate.py --model Qwen/Qwen3.5-0.8B \
    --adapter <path>/netoracle-0.8b/seed41 \
    --prompts dataset/ft/eval_prompts.jsonl --out gen.jsonl

The gated policy in the paper samples 5 answers at temperature 0.8 and changes a knob only if at least 4 agree on a gain of at least 15%.

Limitations

• Trained only on an emulated testbed (network namespaces, HTB bottleneck, netem delay and loss) on one kernel version.
• On real paths whose bottleneck the host cannot observe (the Pantheon archive), the emulation-only model barely beats always predicting no change. Training on measurements from other real paths raises direction accuracy to 0.68; those replay models are not part of this release.
• One short measurement cannot predict tail latency dominated by retransmission timeouts.

License

The adapters are released under the MIT license. Use of the base models is governed by their own licenses.

Citation

@inproceedings{netoracle2026,
  author    = {Luyao Bai and Xiaoguang Wang},
  title     = {Toward a World Model for the Linux Kernel Network Stack},
  booktitle = {Proceedings of the 25th ACM Workshop on Hot Topics in Networks (HotNets '26)},
  year      = {2026},
  address   = {Salt Lake City, UT, USA},
  publisher = {ACM},
  doi       = {10.1145/3847290.3859973}
}