duccio
@duccio.bsky.social
Reposted by duccio
🚨 First ever post here~ New paper out! Instruct models are not always the best. Scaling down 📉 instruction tuning strength via partial adaptation leads to material gains 🚀 on few-shot in-context learning NLP tasks across model families and sizes.
arxiv.org/abs/2504.11626
arxiv.org/abs/2504.11626
April 17, 2025 at 8:55 PM
🚨 First ever post here~ New paper out! Instruct models are not always the best. Scaling down 📉 instruction tuning strength via partial adaptation leads to material gains 🚀 on few-shot in-context learning NLP tasks across model families and sizes.
arxiv.org/abs/2504.11626
arxiv.org/abs/2504.11626