paraphrasing — DeepSeek-R1-Distill-Qwen-7B + LoRA
Reasoning-model fine-tune for paraphrase generation. Multi-domain LoRA training across literary, technical, and academic text. BERTScore 0.9521 beats BART and T5 on the same evaluation set.
What this is
LoRA fine-tune of DeepSeek-R1-Distill-Qwen-7B for paraphrasing. The model keeps its chain-of-thought during training, so it reasons before producing a paraphrase.
How it works
Built a multi-domain training set: literary fiction, technical docs (React, TensorFlow), academic papers, and news. Generated gold paraphrases with OpenAI as SFT targets. Trained with LoRA to keep the parameter count small.
Results
BERTScore 0.9521, BLEU diversity 0.5135, edit distance 0.3439, syntactic diversity 0.1472, harmonic 0.4684. Beats BART and T5 on the same eval set.