SmolLM2-1.7B-Instruct
HuggingFaceTB/SmolLM2-1.7B-Instruct · Apache-2.0
A 1.7B model good enough to be your default on-device LLM.
この解説はまだ日本語がありません。英語の原文を表示しています。
SmolLM2 is the small-language-model sweet spot: strong enough for summarisation, extraction and chat, yet light enough to run on a phone or a CPU-only server.
スコア
導入の要点
pip install transformers 深読み
SmolLM2-1.7B-Instruct is a 1.7B instruction-tuned language model that punches well above its weight — the mid-size in the SmolLM2 family, trained on a carefully filtered mixture including UltraFeedback and Argilla synthetic data.
The architecture is a standard transformer decoder with long-context support, and the model card documents the system and instruct prompt formats that work best with it.
Because it is small, it is a first-choice model for on-device experiments, fine-tuning practice and transformers.js browser demos — and the same weights serve as a teaching reference for the whole SmolLM2 family.