DeepSeek-R1-8B + LoRA + NEFTune Hits 0.901 F1 on Financial NER
June 10, 2026
Fine-tuning DeepSeek-R1-8B with LoRA and NEFTune on 1,693 instruction-formatted financial sentences achieves micro-F1 of 0.901 across seven entity types including Company, Money, and Product. NEFTune's uniform embedding noise provides a measurable boost over LoRA alone on this domain-specific task.
HOW THIS AFFECTS YOU
●
builderThe LoRA + NEFTune recipe on a small labeled corpus is a low-cost path to strong financial NER if you're building knowledge graph extraction pipelines from financial documents.