Automated Review of Academic Literature Generation and Assessment Dataset

Literature abstracts in Chinese Peking University core journals and English SCI journals are selected as the data source, and step-level abstracts under the same subject term are obtained through step recognition and vector search, and the step-level literature review dataset is generated based on multiple abstracts using a big language model with manual review and revision for fine-tuning the local big language model. 2. results were obtained using the baseline models ChatGLM3-6B and GLM-3-Turbo, GPT-3.5-turbo, and the fine-tuned model GLM-Lora for simultaneous literature review generation of the validation set. 3. using the manually reviewed original review as the reference text, the generated results of different models were evaluated using TF-IDF weighted cosine similarity, BLEU, ROUGE metrics. 4. large model assessment of the generated results of different models using GEMINI-Pro, GPT-4 and also GPT-4 for authenticity assessment.