README.md

---
language:
  - zh
license: apache-2.0
tags:
  - finance
  - cryptocurrency
  - chinese
  - news-scoring
  - text-classification
  - text-regression
pipeline_tag: text-classification
library_name: transformers
base_model: LocalOptimum/chinese-crypto-sentiment
metrics:
  - mae
  - accuracy
  - pearsonr
model-index:
  - name: chinese-crypto-importance (v1.1)
    results:
      - task:
          type: text-classification
          name: News Importance Binning
        metrics:
          - type: mae
            value: 6.87
            name: MAE
          - type: accuracy
            value: 61.8%
            name: Bin Accuracy
          - type: pearsonr
            value: 0.532
            name: Pearson r
---

# Chinese Crypto News Importance Scoring Model | 中文加密货币新闻重要性评分模型 (v1.1)

## 模型描述 | Model Description

本模型基于 [LocalOptimum/chinese-crypto-sentiment](https://huggingface.co/LocalOptimum/chinese-crypto-sentiment) 进行 LoRA 微调，专门用于评估中文加密货币新闻的“市场重要性”，而不是传统的情感极性。

模型采用双头结构，同时输出：

- `importance_score`：0-100 连续分数，用于衡量新闻对市场的潜在影响
- `importance_bin`：4 档区间分类，分别为 `noise` / `low` / `medium` / `high`

它要回答的问题是：这条新闻是否值得交易员、研究员或自动化新闻流优先关注，而不只是判断文本是利好还是利空。

This model is LoRA fine-tuned from [LocalOptimum/chinese-crypto-sentiment](https://huggingface.co/LocalOptimum/chinese-crypto-sentiment) for Chinese cryptocurrency news importance scoring rather than plain sentiment classification. It outputs both a continuous score and a 4-way importance bin for ranking and filtering workflows.

## 训练数据 | Training Data

- 数据量 | Size: 20286 条中文加密货币新闻样本 | 20286 Chinese crypto news samples
- 数据来源 | Source: EventAlpha / WatchTower 采集的 19729 条新闻 + 557 条推文 | 19729 news articles + 557 tweets collected via EventAlpha / WatchTower
- 标注方式 | Labeling: 自动四维评分管线 + 规则修正 | 4-axis automatic scoring pipeline with rule-based cleanup
- 划分方式 | Split: 随机划分，训练集 17243 / 验证集 3043 | Random split with 17243 train and 3043 validation samples
- 平均分数 | Average Score: 41.7

### 标注维度 | Scoring Axes

| Axis | Range | Description |
|---|---:|---|
| Market Reaction | 0-40 | Post-news price move, volume expansion, and volatility reaction |
| Novelty | 0-30 | Whether the item is first-hand, repeated, or part of a digest |
| Content Quality | 0-20 | Information density, numeric detail, token relevance, and noise penalties |
| Source Authority | 0-10 | Credibility of the outlet, platform, and whether it is official |

### 数据分布 | Label Distribution

| Bin | Score Range | Count | Share | 含义 / Interpretation |
|---|---:|---:|---:|---|
| `noise` | 0-25 | 1626 | 8.0% | Low-signal, duplicate, digest, or weakly relevant content |
| `low` | 25-50 | 14773 | 72.8% | Routine updates that rarely move the market on their own |
| `medium` | 50-75 | 3840 | 18.9% | Tradeable developments with meaningful but limited impact |
| `high` | 75-100 | 47 | 0.2% | Major events that may materially change price or risk appetite |

## 性能指标 | Performance Metrics

当前公开版本在验证集上的表现如下：

| 指标 Metric | 数值 Value |
|---|---:|
| MAE | 6.87 |
| Bin Accuracy | 61.8% |
| Pearson r | 0.532 |
| Best Epoch | 4 |

## 分数解释 | Score Interpretation

| Bin | Score Range | 典型含义 |
|---|---:|---|
| `noise` | 0-25 | 摘要类、弱相关信息、重复快讯、低信号内容 |
| `low` | 25-50 | 常规更新、普通运营动作、主观评论、有限催化 |
| `medium` | 50-75 | 有交易意义的重要进展，但未必足以改变大趋势 |
| `high` | 75-100 | 黑客攻击、ETF 获批、重大监管变化、系统性风险事件 |

## 使用方法 | Usage

### 方式一：加载完整双头模型（推荐） | Option 1: load the full dual-head model

这种方式可以同时得到 `importance_score` 和 `importance_bin`。

```python
import __main__
import sys
import torch
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

repo_id = "LocalOptimum/chinese-crypto-importance"
local_dir = snapshot_download(repo_id)
sys.path.insert(0, local_dir)

from model import NewsImportanceModel

__main__.NewsImportanceModel = NewsImportanceModel

tokenizer = AutoTokenizer.from_pretrained(local_dir)
model = torch.load(f"{local_dir}/model.pt", map_location="cpu", weights_only=False)
model.eval()

text = "美国现货以太坊 ETF 获批"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits, score = model(
        input_ids=inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        token_type_ids=inputs.get("token_type_ids"),
    )
    probs = torch.softmax(logits, dim=-1)[0]
    labels = ["noise", "low", "medium", "high"]
    importance_bin = labels[probs.argmax().item()]
    importance_score = score.item() * 100

print(importance_bin)
print(round(importance_score, 1))
```

### 方式二：仅使用 HuggingFace 分类头 | Option 2: use the classification head only

这种方式兼容 `pipeline("text-classification")`，但只能直接输出 4 档分类，不包含连续分数。

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

repo_id = "LocalOptimum/chinese-crypto-importance"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

pipe = pipeline("text-classification", model=model, tokenizer=tokenizer)
print(pipe("比特币突破关键阻力位并创下阶段新高"))
```

## 训练配置 | Training Configuration

- 基础模型 | Base Model: `LocalOptimum/chinese-crypto-sentiment`
- 模型结构 | Architecture: BERT backbone + classification head + regression head
- 最大长度 | Max Length: 256
- 训练轮数 | Epochs: 10（Early Stopping patience=3，最佳 epoch=4）
- 批次大小 | Batch Size: 16
- 学习率 | Learning Rate: 2e-5
- LoRA: `r=16`, `alpha=32`, `dropout=0.05`
- 损失函数 | Loss: `0.6 * cross_entropy + 0.4 * mse`
- 混合精度 | Mixed Precision: FP16

## 适用场景 | Use Cases

- 加密货币新闻优先级排序
- 实时快讯过滤与告警降噪
- 研究员 / 交易员新闻流预筛选
- 回测与研究中的事件权重特征构建
- 市场重大事件回溯分析

## 核心标注原则 | Annotation Principles

- 重要性不等于情绪：利好和利空都可能是高重要性
- 优先看市场反应，再结合新颖度、内容质量和来源可信度
- 重复快讯、摘要汇总、弱相关宏观噪声会被系统性降分
- 官方公告、重大安全事件、ETF / 监管突破通常更高分
- 主观观点和常规运营更新通常落在 `low` 或 `noise`

## 局限性 | Limitations

- 数据分布明显偏向 `low`，当前版本对高重要性事件仍偏保守
- `high` 样本较少，模型对极端高分事件的区分能力仍有提升空间
- 主要适用于中文加密货币新闻，跨领域泛化能力有限
- HuggingFace 原生 `pipeline` 只暴露分类头；连续分数需要加载 `model.pt`
- 标签来自自动评分管线与规则修正，不等同于大规模人工金融标注

## 许可证 | License

Apache-2.0

## 引用 | Citation

如果你在研究或产品中使用本模型，可以引用：

```bibtex
@misc{onefly_crypto_importance_2026,
  title={Chinese Crypto News Importance Scoring Model},
  author={Onefly},
  year={2026},
  howpublished={\url{https://huggingface.co/LocalOptimum/chinese-crypto-importance}},
  note={LoRA fine-tuned from LocalOptimum/chinese-crypto-sentiment, 20286 samples, MAE=6.87, BinAcc=61.8%}
}
```

## 基础模型 | Base Model

本模型基于以下模型继续训练：

- [LocalOptimum/chinese-crypto-sentiment](https://huggingface.co/LocalOptimum/chinese-crypto-sentiment)

## 更新日志 | Changelog

### 当前公开版本 | Current Public Version

- 首个公开的重要性评分模型版本
- 支持双头输出：连续重要性分数 + 4 档重要性分类
- 基于 20286 条中文加密货币新闻样本完成训练
- 当前验证指标：MAE=6.87，Bin Accuracy=61.8%，Pearson r=0.532

如有问题或建议，欢迎提 issue 或 PR。