laya-multilingual-UNSW-NB15

Supervised post-training of convaiinnovations/laya-multilingual for binary Normal / Attack classification of UNSW-NB15 network flows. This is a Laya decision model, used through laya, rather than a text generation model.

Training

The completed local laya_posttraining_v1 run used Laya 0.4.7 and all 42 flow features. IDs, labels and attack categories are excluded from model input. A 4,000-row training input supplied 3,600 gradient-training items and 400 temperature-calibration items. Training used four epochs, hard labels with cross-entropy (soft-ce), full encoder fine-tuning, micro-batch 1, gradient accumulation 16, encoder learning rate 2.5e-5, head learning rate 1e-4, and seed 2026. Maximum input length is 2,048 tokens; head maximum length is 128. The final epoch checkpoint is published.

Training preparation removed exact feature overlaps with the official test set, conflicting labels and duplicate training flows. Separate cutoff and validation splits came from training data. Base revision: 1720e3e3357cfe1e281542e223f8273b0890ca34.

Loading

pip install laya==0.4.7 huggingface_hub

import json
from pathlib import Path
import laya
from huggingface_hub import snapshot_download

path = Path(snapshot_download("anguswwww/laya-multilingual-UNSW-NB15"))
agent = laya.load(str(path), device="cpu")
questions = json.loads((path / "questions.json").read_text())
# serialized_features must contain all 42 features in features.json order,
# as name=value pairs separated by ';' (no added spaces). Use the original numeric precision.
flow_text = "UNSW-NB15 network flow. Values describe a connection.\nTARGET: " + serialized_features
answer = agent.system_one(flow_text, questions, max_len=2048, head_max_len=128)
print(answer["answers"])

Use the original flow serializer and reject truncated inputs. features.json records feature order. The frozen fine-tuned binary threshold is 0.1728 on the Attack probability, chosen for F2 on the cutoff split; see frozen.json. This threshold is separate from Laya's internal abstention threshold. Preserve the calibrated configuration when loading.

Evaluation and limitations

test_metrics.csv contains the original separate 2,000-flow held-out comparison, and validation_metrics.csv contains development results. These files are not the later matched five-run paper evaluation. The training report's 400-item calibration results are in-sample calibration metrics, not independent held-out performance.

This is a research checkpoint specialized for the UNSW-NB15 feature representation. Generalization to other traffic, feature extraction systems, prevalence, or attack distributions is not established. The F2 threshold prioritizes recall and is not an operational cost model. The multilingual base does not establish multilingual intrusion-detection performance after this training.

Files

The repository includes final model weights, encoder configuration, tokenizer, calibrated Laya configuration, questions, feature order, training reports, frozen thresholds and evaluation summaries. Intermediate checkpoints and raw training/test rows are excluded.