QLoRA 微调用于 JSON 结构化提取实战

dev.to · AI#QLoRA#微调#LLM

Editor brief · 编辑速览

在 Qwen2.5-1.5B-Instruct 模型上分别用 4/8/16 比特量化训练，评估 Exact Match 和 Field Match 指标，为需要私有化部署结构化数据提取能力的团队提供量化精度参考。

文章思维导图

Knowledge map

拖拽缩放

Full translation

完整中文译文

我们需要从一个领域特定的非结构化文本中提取结构化的 JSON。决定分别用 4-bit、8-bit 和 16-bit 三种量化级别对 Qwen2.5-1.5B-Instruct 模型进行训练。

数据集总样本数：约 280

训练使用 QLoRA 微调

我们按每个数据集分割测量两个分数：

Exact Match（精确匹配）：每样本二元分数（所有字段完全匹配为 1，否则为 0）。在分割数据集上汇总。

Field Match（字段匹配）：每样本的部分得分（正确字段数 / 总字段数）。在分割数据集上汇总。

标准 JSON

{
  "first_name": "Jordan",
  "last_name": "Kaur",
  "employment_type": "contractor",
  "employer_name": "Grounded Movement",
  "employer_email": "payroll@goat.com",
  "employer_address_line_1": "41 King Street",
  "employer_address_line_2": "Madison, UK",
  "employer_registration_number": null,
  "employer_tax_number": null,
  "currency": "USD",
  "pay_rate": 45.0,
  "pay_rate_basis": "per_shift",
  "incentive_pay": 5.0,
  "incentive_type": "per_reservation_over_n",
  "n": 15,
  "reservation_types": [
    "check_ins"
  ],
  "exclude_staff_reservations": true,
  "exclude_cancelled_shifts": true,
  "max_pay": null,
  "payment_frequency": "weekly"
}

预测 JSON（模型输出）

{
  "first_name": "Jordan",
  "last_name": "Kaur",
  "employment_type": "contractor",
>>"employer_name": "Grounded",<<
  "employer_email": "payroll@goat.com",
  "employer_address_line_1": "41 King Street",
  "employer_address_line_2": "Madison, UK",
  "employer_registration_number": null,
  "employer_tax_number": null,
  "currency": "USD",
  "pay_rate": 45.0,
>>"pay_rate_basis": null,<<
  "incentive_pay": 5.0,
  "incentive_type": "per_reservation_over_n",
  "n": 15,
  "reservation_types": [
    "check_ins"
  ],
  "exclude_staff_reservations": true,
  "exclude_cancelled_shifts": true,
  "max_pay": null,
  "payment_frequency": "weekly"
}

Mismatch 1: "employer_name": "Grounded" (Predicted) vs "Grounded Movement" (Gold)
Mismatch 2: "pay_rate_basis": null (Predicted) vs "per_shift" (Gold)

Total Fields = 20
Correct Fields = 18
Wrong Fields = 2

Exact Match
Score: 0/1

Field Level Match
Score: 18/20 = 0.9  (90%)

虽然我们用 4-bit、8-bit 和 16-bit QLoRA 进行了实验，但为简洁起见，低位数的实验结果在此省略。

-------------------------------------------------------------------
Model                    Exact Accuracy    Field Match
-------------------------------------------------------------------
Baseline                 0.00              0.54
Fine‑tuned               0.62              0.97

微调后字段匹配从约 54% 提升至约 97%。精确匹配虽然较低（62%），但从 0% 实现了零的突破。

字段级结果表明，adapter 成功学习到了大多数字段的正确键值映射。