---

language:

• en license: apache-2.0 library_name: transformers pipeline_tag: text-classification tags:
• llm-detection
• ai-generated-text-detection
• academic-writing
• peer-review
• qwen
• lora
• binary-classification base_model: Qwen/Qwen3-0.6B-Base model_type: qwen3

---

Qwen3-0.6B Academic LLM-Writing Detector

Model Summary

This model is a parameter-efficient LoRA fine-tuning of Qwen3-0.6B-Base for binary classification of academic text. It is designed to identify whether a text segment is more likely to be human-written or LLM-assisted / LLM-like.

The model was developed for research on AI-mediated scholarly communication, with downstream applications including the classification of:

• article abstracts,
• introduction sections,
• peer-review reports.

The classifier outputs:

• 0: human-written / human-like
• 1: LLM-assisted / LLM-like

Intended Use

Primary intended use

This model is intended for:

• research on the diffusion of LLM use in academic writing,
• large-scale measurement of LLM-associated language patterns in scholarly corpora,
• comparative analyses across article sections, journals, or peer-review texts.

Out-of-scope use

This model is not intended for:

• plagiarism adjudication,
• misconduct accusations against individual authors or reviewers,
• legal, institutional, or disciplinary enforcement,
• high-stakes decisions about individuals without additional human review.

Predictions should be interpreted as probabilistic indicators of LLM-associated writing style, not as definitive proof of AI use.

Training Data

The model was fine-tuned on the PAN25 Generative AI Detection Task 1 training data, using a binary text-classification objective.

The task setting is:

• detect whether a text is human-written or generated / AI-assisted,
• optimized as a supervised binary classification problem.

Training Procedure

Base model

• Qwen3-0.6B-Base

Fine-tuning method

• LoRA fine-tuning for sequence classification

Task

• Binary sequence classification

Label definition

• 0: human-written / human-like
• 1: LLM-assisted / LLM-like

Best checkpoint

• checkpoint-1000
  or, if you prefer the merged release wording:
• released checkpoint: final_best

Evaluation

On the validation split used during fine-tuning, the best checkpoint achieved:

• Accuracy: 0.9980
• Precision: 1.0000
• Recall: 0.9970
• F1: 0.9985

More specifically:

{
  "eval_loss": 0.007584796752780676,
  "eval_accuracy": 0.9980495959877403,
  "eval_precision": 1.0,
  "eval_recall": 0.9969723183391004,
  "eval_f1": 0.99848386398094,
  "eval_tp": 2305.0,
  "eval_tn": 1277.0,
  "eval_fp": 0.0,
  "eval_fn": 7.0
}