Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment
Democratizing AI with Small Language Models: Structured Benchmarking and Parameter-Efficient Fine-Tuning for Local Deployment
通过小型语言模型实现 AI 民主化：结构化基准测试与本地部署的参数高效微调
Abstract: AI democratization is not primarily a question of matching frontier-scale generality; it is a question of whether capable models can be selected, audited, and specialized under hardware and governance constraints that ordinary institutions can actually satisfy. 摘要： AI 的民主化主要不是为了追求与前沿规模模型同等的通用性，而是探讨在普通机构能够满足的硬件和治理约束下，是否能够筛选、审计并专门化处理能力强的模型。
This paper studies that problem through a controlled evaluation of nine open-weight language models between 135M and 3B parameters on a 1,085-example, 16-topic multiple-choice benchmark designed for structured local deployment. 本文通过对九种参数量在 1.35 亿至 30 亿之间的开源权重语言模型进行受控评估，研究了这一问题。评估基于一个包含 1,085 个样本、涵盖 16 个主题的多项选择基准测试，该基准专为结构化本地部署而设计。
The benchmark emphasizes symbolic precision, constrained formatting, extraction, and short-horizon semantic decision making under a strict one-letter output protocol. 该基准测试强调在严格的单字母输出协议下，模型在符号精度、约束格式化、信息提取以及短视界语义决策方面的表现。
A shared parameter-efficient fine-tuning pipeline then adapts a subset of models using 4-bit NF4 quantization with DoRA/LoRA-style adapters on an NVIDIA L4-class budget. 随后，研究采用统一的参数高效微调流程，在 NVIDIA L4 级别的预算限制下，利用 4-bit NF4 量化和 DoRA/LoRA 风格的适配器对部分模型进行了适配。
In base evaluation, Qwen Coder 3B leads at 75.67% strict accuracy, followed by Qwen2.5 1.5B at 67.10%, Qwen3.5 2B at 64.98%, and Granite 3.3 2B at 64.61%. 在基础评估中，Qwen Coder 3B 以 75.67% 的严格准确率领先，紧随其后的是 Qwen2.5 1.5B（67.10%）、Qwen3.5 2B（64.98%）和 Granite 3.3 2B（64.61%）。
On the shared 108-example held-out fine-tuning split, adaptation improves Qwen Coder 3B by +26.85 points, SmolLM2 1.7B by +25.92, Qwen2.5 1.5B by +19.44, SmolLM2 360M by +10.18, and SmolLM2 135M by +5.55. 在 108 个样本的留出微调集上，经过适配后，Qwen Coder 3B 的准确率提升了 26.85 个百分点，SmolLM2 1.7B 提升了 25.92 个百分点，Qwen2.5 1.5B 提升了 19.44 个百分点，SmolLM2 360M 提升了 10.18 个百分点，SmolLM2 135M 提升了 5.55 个百分点。
Across ranking, topic-level heterogeneity, difficulty strata, failure composition, efficiency frontiers, and topic-conditioned transfer, the same conclusion recurs: a disciplined workflow of benchmark construction, cross-model evaluation, and low-cost specialization already makes a subset of sub-3B models viable as local experts for structured niche workloads. 通过对排名、主题异质性、难度分层、失败构成、效率边界以及主题条件迁移的综合分析，得出了一个反复出现的结论：通过规范化的基准构建、跨模型评估和低成本专门化工作流，一部分 30 亿参数以下的小型模型已经能够胜任结构化利基工作负载，成为可行的本地专家。