Researchers have developed DistilledGemma, an efficient system for extracting person-place relationships from multilingual historical articles, achieving a 0.688 mean score in the HIPE-2026 shared task. The system employs a three-stage knowledge distillation pipeline, starting with prompt engineering across large language models, followed by supervised fine-tuning of a Gemma 4 26B model using QLoRA, and concluding with response-level distillation into a smaller Gemma 4 E2B student model. This approach successfully reduced model size while maintaining strong reasoning capabilities, ranking second in efficiency-accuracy profiles for both standard and binary test sets. AI
IMPACT Demonstrates effective knowledge distillation for efficient processing of historical documents, potentially lowering computational costs for similar NLP tasks.
RANK_REASON The cluster describes a research paper detailing a new model and methodology for relation extraction, including performance metrics.
Read on Hugging Face Daily Papers →
- English
- French
- German
- HIPE-2020
- HIPE-2022
- HIPE-2026
- large-language models
- Multilingual Historical Texts
- Person-Place Relation Extraction
- DistilledGemma
- Gemma 4 26B-A4B
- Gemma 4 E2B
- QLoRA
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
