YiMeng-SYSU/roberta-logic-sentiment-zh

RoBERTa-Logic-Finetuned on Chinese Logic Dataset

->中文

This model is a RoBERTa model fine-tuned on a Chinese logic dataset to perform logic-related sentiment classification tasks. The dataset was generated using the Doubao API. It achieves an F1 score of 96.00% on the validation set.

Model Details

• Architecture: RoBERTa
• Base Model: uer/roberta-base-finetuned-dianping-chinese
• Framework: Transformers, PyTorch
• Fine-tuning Dataset: YiMeng-SYSU/chinese-logic-sentiment-dataset (Generated by Doubao API)
• Hardware: Trained on NVIDIA RTX 5070 Ti + AMD 9800X3D

Performance

The performance of this model on the Golden Test Set (960 samples) is as follows:

Metric	Score
Accuracy	96.15%
Macro-F1	96.00%
Recall (Negative)	97.53%
Recall (Positive)	94.15%

Detailed Performance by Logic Type

Type	Accuracy	Note
Simple (简单句)	97.50%	基础稳固
Transition (转折)	97.03%	完美理解转折逻辑
Double Neg (双重否定)	96.67%	彻底解决"双重否定表肯定/否定"的歧义
Irony (反讽)	92.47%	核心亮点：具备识别"反话正说"的能力

Usage

You can use the model for sentiment analysis tasks in Chinese by loading it through the Hugging Face Transformers library.

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch
model_name = "YiMeng-SYSU/roberta-logic-sentiment-zh"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
text = "这部剧简直有毒，害得我昨晚又熬到凌晨三点，黑眼圈都出来了!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)  # 补充tokenizer参数，更健壮
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
# 结合config中的id2label映射，输出可读标签（而非数字）
id2label = {0: "Negative (负面)", 1: "Positive (正面)"}
pred_label_id = torch.argmax(probs).item()
print(f"Label: {id2label[pred_label_id]}")
print(f"Confidence: {probs.max().item():.4f}")