基于 LlamaFactory 微调大模型的实体识别的评估实现

{
    "prompt": "从以下政府文本中，提取指定的实体类别，并按照JSON格式输出结果。  xxx",      
    "predict": "\n```json\n{\n  \"行政区划和地理实体\": [],\n  \"政府机构和组织\": [],\n ... \"特定术语和关键词\": [\"改革创新\", \"体制机制障碍\", \"行政管理体制\", \"城乡一体化发展机制\"]\n}\n```", 
    "label": "```json\n{\n  \"行政区划和地理实体\": [],\n  \"政府机构和组织\": [], ... \"特定术语和关键词\": [\"改革创新发展\", \"高质量发展\", \"高品质生活\", \"体制机制障碍\"]\n}\n```"
    ...
}
1.
2.
3.
4.
5.
6.

import os
import re
import json
from dataclasses import dataclass

@dataclass
class EvalNode:
    # 默认值
    predict_right_num: int = 0
    predict_num: int = 0
    label_num: int = 0

    @property
    def accuracy(self):
        return self.predict_right_num / (self.predict_num + 1e-6)

    @property
    def recall(self):
        return self.predict_right_num / (self.label_num + 1e-6)

    @property
    def f1(self):
        return 2 * self.accuracy * self.recall / (self.accuracy + self.recall + 1e-6)

@dataclass
class NEREvaluate:
    ent_class: list[str]
    input_file: str
    avg_accuracy: int = 0
    avg_recall: int = 0
    avg_f1: int = 0
    predict_error: int = 0

    def __post_init__(self):
        self._evaluate_by_jsonl()

    def _evaluate_by_jsonl(self):

        with open(self.input_file, "r", encoding="utf-8") as f:
            self.total_ent = {ent: EvalNode() for ent in self.ent_class}

            for line in f:
                data = json.loads(line)

                # 大模型采取的是序列到序列到文本生成，不能转换为字典的数据跳过即可
                try:
                    matches = re.search("\{.*?\}", data["predict"], re.DOTALL)
                    if matches:
                        predict = eval(matches.group(0))
                except:
                    self.predict_error += 1
                    predict = {ent: [] for ent in self.ent_class}

                try:
                    # 防止label出错
                    matches = re.search("\{.*?\}", data["label"], re.DOTALL)
                    if matches:
                        labels = eval(matches.group(0))
                except:
                    continue

                # 每个不同的实体类别单独计数
                for ent_name in self.ent_class:
                    label_set = set(labels.get(ent_name, []))
                    predict_set = set(predict.get(ent_name, []))
                    self.total_ent[ent_name].predict_right_num += len(
                        label_set & predict_set
                    )
                    self.total_ent[ent_name].predict_num += len(predict_set)
                    self.total_ent[ent_name].label_num += len(label_set)

        for ent in self.ent_class:
            self.avg_accuracy += self.total_ent[ent].accuracy
            self.avg_recall += self.total_ent[ent].recall
            self.avg_f1 += self.total_ent[ent].f1

        self.avg_accuracy = self.avg_accuracy / len(self.ent_class)
        self.avg_recall = self.avg_recall / len(self.ent_class)
        self.avg_f1 = self.avg_f1 / len(self.ent_class)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.

ENT_CLASS = [
    "行政区划和地理实体",
    "政府机构和组织",
    "法律法规与政策",
    "人物与职务",
    "产业",
    "产业园区",
    "高新技术",
    "项目与工程",
    "社会服务与基础设施",
    "特定术语和关键词",
]
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

ner_eval = NEREvaluate(
        ENT_CLASS, "xxx/generated_predictions.jsonl")
    )

for key, item in ner_eval.total_ent.items():
    print(key, item.f1)
1.
2.
3.
4.
5.
6.

行政区划和地理实体 0.6292412411907403
人物与职务 0.45714235316380664
...
特定术语和关键词 0.29256472295174
1.
2.
3.
4.