traffic-classify

Chinese text classifier for traffic / intent classification into three labels:

• 0: 非研发相关
• 1: 研发相关
• 2: 中性

Files

• Root directory: Transformers model, tokenizer, and config.
• onnx/model_fp32.onnx: FP32 ONNX export.
• onnx/model_int8.onnx: INT8 quantized ONNX model for CPU inference.

Evaluation

On the held-out evaluation set, the classifier achieves an overall accuracy of approximately 94%, with a weighted F1-score of approximately 94%, demonstrating strong classification capability across different traffic types.

The per-class F1-scores are:

• 研发相关: 0.96
• 中性: 0.93
• 非研发相关: 0.88

The results indicate that the model performs particularly well in identifying 研发-related traffic, while non-研发 traffic remains relatively more challenging due to semantic overlap with technical discussions.

Dataset Policy

The dataset uses a compact three-class policy:

• 研发相关: full software / IT / electronics / communication terminology sources.
• 中性: stopwords only.
• 非研发相关: remaining dictionary categories, compactly sampled.

See the project DATA_SOURCES.md for source details.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

repo_id = "pawizard/traffic-classify"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

text = "这个 NullPointerException 报错怎么修复"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    pred = int(logits.argmax(dim=-1).item())

print(model.config.id2label[pred])