DocuBrief — Document Summarization Assistant (fine-tuned Qwen/Qwen2.5-1.5B-Instruct)
Fine-tuned with QLoRA on 4000 examples derived from CNN/DailyMail.
Intended use
Designed to be paired with a RAG/chunking layer that supplies the actual document content at inference time (see the DocuBrief AI project guide). This model provides the summarization behavior - clear, well-structured, faithful summaries - not memorized knowledge of any particular document.
Training
- Base model:
Qwen/Qwen2.5-1.5B-Instruct - Method: QLoRA (4-bit NF4), r=16, alpha=32
- Steps: 200, LR: 2e-4, max seq len: 1024
- Dataset: CNN/DailyMail 3.0.0, articles truncated to <= 500 words
Limitations
- English only. Small model - pair with the map-reduce/RAG layer described in the DocuBrief AI project guide for accurate summarization of large or arbitrary documents.
- Trained on news-style text; may need a larger base model or more training for highly technical or domain-specific documents (legal, medical, scientific).