Chinese Financial Sentiment Analysis Model (Crypto Focus)

中文金融情感分析模型(加密货币领域)

模型描述 | Model Description

本模型基于 yiyanghkust/finbert-tone-chinese 经过多轮迭代微调,专门用于分析中文加密货币相关新闻和社交媒体内容的情感倾向。模型可以识别三种情感类别:正面(Positive)、中性(Neutral)和负面(Negative)。

训练数据经过 Claude AI 逐条人工审阅、纠正标注错误,确保数据质量。

This model is iteratively fine-tuned from yiyanghkust/finbert-tone-chinese, specifically designed for sentiment analysis of Chinese cryptocurrency-related news and social media content. It classifies text into three sentiment categories: Positive, Neutral, and Negative.

Training data is manually reviewed and corrected entry-by-entry by Claude AI to ensure annotation quality.

训练数据 | Training Data

  • 数据量 | Size: 2208条人工审阅标注的中文金融新闻 | 2208 manually reviewed Chinese financial news articles
  • 数据来源 | Source: 加密货币相关新闻和推文 | Cryptocurrency-related news and tweets
  • 标注方式 | Annotation: 模型预测 + Claude AI 逐条审阅纠正 | Model prediction + Claude AI manual review & correction
  • 数据分布 | Distribution:
    • Positive(正面): 734条 (33.2%)
    • Neutral(中性): 899条 (40.7%)
    • Negative(负面): 575条 (26.0%)

性能指标 | Performance Metrics

在442条测试集上的表现(80/20分层划分) | Performance on 442 test samples (80/20 stratified split):

指标 Metric 数值 Value
准确率 Accuracy 84.84%
F1分数 F1 Score 84.88%
精确率 Precision 85.36%
召回率 Recall 84.84%

各类别详细指标 | Per-class Metrics

类别 Class Precision Recall F1
negative 0.938 0.791 0.858
neutral 0.806 0.878 0.840
positive 0.846 0.857 0.851
weighted avg 0.854 0.848 0.849

性能迭代历史 | Performance History

版本 Version 训练数据 Data F1 Score Accuracy
v1.0 500条 61.65%
v2.0 1000条 63.65% 64.50%
v3.5 1500条 67.16% 68.33%
v4.0 1700条 70.91% 72.06%
v5.0 2008条 76.88% 77.36%
v6.0 2208条 84.88% 84.84%

使用方法 | Usage

快速开始 | Quick Start

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# 加载模型和分词器 | Load model and tokenizer
model_name = "LocalOptimum/chinese-crypto-sentiment"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# 分析文本 | Analyze text
text = "比特币突破10万美元创历史新高"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

# 预测 | Predict
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(predictions, dim=-1).item()

# 结果映射 | Result mapping
labels = ['positive', 'neutral', 'negative']
sentiment = labels[predicted_class]
confidence = predictions[0][predicted_class].item()

print(f"情感: {sentiment}")
print(f"置信度: {confidence:.4f}")

批量处理 | Batch Processing

texts = [
    "币安获得阿布扎比监管授权",
    "以太坊完成Fusaka升级",
    "某交易所遭攻击损失100万美元"
]

inputs = tokenizer(texts, return_tensors="pt", truncation=True,
                   max_length=128, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_classes = torch.argmax(predictions, dim=-1)

labels = ['positive', 'neutral', 'negative']
for text, pred in zip(texts, predicted_classes):
    print(f"{text} -> {labels[pred]}")

训练参数 | Training Configuration

  • 基础模型 | Base Model: yiyanghkust/finbert-tone-chinese(经多轮迭代微调)
  • 训练轮数 | Epochs: 5(Early Stopping patience=3,Epoch 5 达到最佳)
  • 批次大小 | Batch Size: 16
  • 学习率 | Learning Rate: 2e-5
  • 最大序列长度 | Max Length: 128
  • 训练设备 | Device: NVIDIA GeForce RTX 5080 Laptop GPU (16GB)
  • 混合精度 | Mixed Precision: FP16
  • 最佳模型选择 | Best Model: metric_for_best_model='f1'

适用场景 | Use Cases

  • ✅ 加密货币新闻情感分析
  • ✅ 社交媒体舆情监控
  • ✅ 金融市场情绪指标
  • ✅ 实时新闻情感跟踪
  • ✅ 投资决策辅助参考

核心标注原则 | Annotation Principles

  • 加密货币是风险资产(类似美股),不是避险资产(类似黄金)
  • 战争、地缘冲突、关税 → negative(利空风险资产)
  • 平台上线新币种/功能 → neutral(常规运营,非利好)
  • 个人观点/分析师预测 → neutral(主观意见)
  • 明确利好(ETF通过、大额买入、政策支持)→ positive
  • 明确利空(清算、暴跌、诈骗、监管打压)→ negative

局限性 | Limitations

  • ⚠️ 主要针对加密货币领域的金融新闻,其他金融领域可能表现不佳
  • ⚠️ 短文本(少于10字)的分析准确率可能下降
  • ⚠️ 仅支持简体中文
  • ⚠️ 模型不能替代人工判断,仅供参考

许可证 | License

Apache-2.0

引用 | Citation

如果使用本模型,请引用:

@misc{watchtower-sentiment-2026,
  title={Chinese Financial Sentiment Analysis Model (Crypto Focus)},
  author={Onefly},
  year={2026},
  howpublished={\url{https://huggingface.co/LocalOptimum/chinese-crypto-sentiment}},
  note={Fine-tuned from yiyanghkust/finbert-tone-chinese, 2208 samples, F1=84.88\%}
}

基础模型 | Base Model

本模型基于以下模型微调:

感谢原作者的贡献!

更新日志 | Changelog

v6.0 (2026-02-28)

  • ✅ 扩充训练数据至2208条(+200条Claude人工审阅数据)
  • ✅ F1分数大幅提升(76.88% → 84.88%,+8.00%)
  • ✅ 大规模纠正地缘政治/战争新闻标注(97条 positive→negative,修复"美以打击伊朗"系统性错误)
  • ✅ negative recall 显著提升(67.0% → 79.1%,+12.1pp)
  • ✅ 地缘政治专项验证:14条测试全部几乎正确(92.9%),8条战争新闻置信度1.00判为negative

v5.0 (2026-02-28)

  • ✅ 扩充训练数据至2008条(+308条Claude人工审阅数据)
  • ✅ F1分数大幅提升(70.91% → 76.88%,+5.97%)
  • ✅ 纠正模型系统性错误(positive→neutral 过度预测等)
  • ✅ 数据分布优化:negative从362增至431条

v4.0 (2026-02-28)

  • ✅ 扩充训练数据至1700条
  • ✅ F1分数提升(67.16% → 70.91%,+3.75%)
  • ✅ 引入Claude AI逐条审阅标注流程

v3.5 (2026-02-27)

  • ✅ 扩充训练数据至1500条
  • ✅ F1分数提升(63.65% → 67.16%,+3.51%)
  • ✅ 大幅修正战争/地缘冲突→positive的系统性错误

v2.0 (2025-12-09)

  • ✅ 扩充训练数据至1000条
  • ✅ 修正标注错误,提升数据质量
  • ✅ F1分数提升(61.65% → 63.65%,+2.01%)

v1.0 (Initial Release)

  • 基于500条标注数据的初始版本

联系方式 | Contact

如有问题或建议,欢迎提 issue 或 PR。


维护者 | Maintainer: Onefly 最后更新 | Last Updated: 2026-02-28

Downloads last month
65
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Evaluation results