[P] QLoRA Fine-Tuning of Qwen2.5-1.5B for CEFR English Proficiency Classification (A1–C2) [P]

2026-05-04 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, short

Summary

A QLoRA fine-tuning project classified English text into 6 CEFR proficiency levels (A1-C2) using the Qwen2.5-1.5B model. The model was fine-tuned with 4-bit NF4 quantization, training only ~0.28% of its parameters. The dataset comprised 1,785 synthetically generated English texts, balanced across CEFR levels and 10 domains, created using the Groq API and Llama-3.3-70B with constraints to preserve linguistic patterns. On a held-out test set of 179 samples, the model achieved 84.9% accuracy and Macro F1 score. Per-level recall ranged from 96.6% for A1 to 60.0% for C2, with most errors occurring between C1 and C2 levels. A FastAPI inference API and Docker deployment setup were also developed.

Key takeaway

For NLP engineers developing language learning applications, this project demonstrates a viable approach to CEFR classification. You should consider QLoRA with smaller models like Qwen2.5-1.5B for efficient deployment. Be aware that synthetic data may introduce distribution shifts, especially for nuanced levels like C2, and plan for validation with authentic learner data to improve real-world performance.

Key insights

QLoRA fine-tuning of Qwen2.5-1.5B effectively classifies CEFR English proficiency, despite synthetic data limitations.

Principles

• Synthetic data can bootstrap NLP tasks.
• QLoRA enables efficient fine-tuning of large models.
• C1/C2 CEFR distinction is challenging for models.

Method

Fine-tune Qwen2.5-1.5B with QLoRA (4-bit NF4) and a linear classification head on a synthetically generated, balanced text dataset for multi-class CEFR proficiency classification.

In practice

• Use QLoRA for memory-efficient fine-tuning.
• Deploy with FastAPI and Docker for inference.
• Consider synthetic data for niche datasets.

Topics

• QLoRA Fine-Tuning
• Qwen2.5-1.5B
• CEFR Classification
• English Proficiency
• Synthetic Data Generation

Best for: Machine Learning Engineer, NLP Engineer, AI Engineer

Related on AIssential

• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific LoRA fine-tuning on Qwen2.5-14B with CCPoetry-49K significantly enhances classical Chinese poetry understanding and translation.
• Building an LLM fine-tuning Dataset — Building multi-speaker Reddit datasets for LLM fine-tuning is complex but achievable, with efficient QLoRA methods requiring minimal samples.
• Fine-Tuning Qwen3.5 — Fine-tuning small vision-language models like Qwen3.5-0.8B on domain-specific datasets significantly improves specialized task performance.
• Small LLMs for Biomedical Claim Verification: Cost-Effective Fine-Tuning, Structural Dataset Shortcuts, and Cross-Domain Generalization — Small, QLoRA-fine-tuned LLMs can surpass larger proprietary models for specialized biomedical claim verification tasks.
• Automated IEP Generation from Traditional Chinese Parent-Teacher Interviews via Corpus-Grounded Feature Diffusion — Corpus-Grounded Feature Diffusion enables high-performance, privacy-preserving automated IEP generation in Traditional Chinese despite low data resources.
• Qwopus and REAP: Custom Qwen3.6 Models for Local Reasoning — Customizing advanced LLMs like Qwen3.6 is difficult, often yielding negative results when considering efficiency and accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.