OniReimu/Edge-Computing-JEV-classifiers

Edge-Computing-JEV service classifiers

Four DistilBERT service classifiers used as reference interpreters in RQ4 (dynamic service catalog) of the paper

Replacing Large Language Models with Jev Decision Models for Low-Latency Edge Service Orchestration Delong Li, Xu Wang, Haochen Gong, Rui Lang, and Guangsheng Yu. University of Technology Sydney.

Each classifier maps a natural-language edge-service request to one service of a fixed catalog (or unsupported). They show what a trained classifier needs when the catalog changes, in contrast to decision models and LLMs that receive the catalog with each request.

Code: github.com/OniReimu/Edge-Computing-JEV · Benchmark and run records: datasets/OniReimu/Edge-Computing-JEV

Models

Every classifier is trained from distilbert/distilbert-base-uncased at revision 12040accade4e8a0f71eabdb258fecc2e7e948be.

Folder	Paper name	Labels	Training examples	Used for
clf_all	DistilBERT-Clf-All	255 (254 services + unsupported)	554: one description per service + 300 development cases of the catalog-size conditions	RQ4 catalog-size conditions (K = 4 to 254)
clf_frozen	DistilBERT-Clf-Frozen	65 (64 services of catalog v0 + unsupported)	322: one description per service + 258 development cases	RQ4 churn conditions, without adaptation
clf_retrained_25	DistilBERT-Clf-Retrained (25% churn)	65 (catalog v1)	339, of which 76 are new labelled examples (16 descriptions of new services + 60 churn development cases)	RQ4 25% churn
clf_retrained_50	DistilBERT-Clf-Retrained (50% churn)	65 (catalog v2)	278, of which 92 are new labelled examples (32 descriptions of new services + 60 churn development cases)	RQ4 50% churn

Training examples come only from the EdgeIntent v1 development split and the catalog descriptions; no test case is used. Hyperparameters are fixed, with no search: max length 128, learning rate 5e-5, batch size 16, 10 epochs, weight decay 0.01, seed 20260924, trained on Apple M4 Max (MPS). Each folder's training.json records the label space, example counts, and training wall time. scripts/eb_rq4_train.py in the GitHub repository rebuilds all four.

Results on the churn conditions

Service top-1 accuracy on the EdgeIntent v1 test split, from experiments/rq1-rq4-interpretation/results/h5_classifier_reference.csv:

Condition	Classifier	Seen services	Unseen services
25% churn	Frozen	0.653	0.000
25% churn	Retrained	0.708	0.147
50% churn	Frozen	0.522	0.000
50% churn	Retrained	0.441	0.142

Results on the catalog-size conditions are in experiments/rq1-rq4-interpretation/results/cells.csv (model DistilBERT-Clf-All).

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo = "OniReimu/Edge-Computing-JEV-classifiers"
tok = AutoTokenizer.from_pretrained(repo, subfolder="clf_all")
model = AutoModelForSequenceClassification.from_pretrained(repo, subfolder="clf_all")
inputs = tok("Please read the licence plate on the gate camera frame, keep it on site.",
             return_tensors="pt", truncation=True, max_length=128)
print(model.config.id2label[model(**inputs).logits.argmax(-1).item()])

Limitations

These are reference baselines trained with a small, fixed recipe on synthetic requests. They are not tuned and are not intended for deployment.

License

Apache-2.0, as the base model.