mags0ft's picture
correct wrong percentage of German Wikipedia pre-training info in README
e409539 verified
---
base_model: HuggingFaceTB/SmolLM2-360M
tags:
- transformers
- unsloth
- llama
license: apache-2.0
language:
- de
datasets:
- wikimedia/wikipedia
- FreedomIntelligence/alpaca-gpt4-deutsch
---
# SmolLM2-360m-German-Instruct
<p align="center">
<img alt="Showcase image for SmolLM2-360m-German-Instruct" src="https://huggingface.co/mags0ft/SmolLM2-360m-German-Instruct/resolve/main/showcase-image.png" width="600" />
</p>
This is a continued pre-train as well as an instruct fine-tune done using Unsloth in order to make SmolLM2 360m capable of speaking German.
It has been trained on 25% of the German Wikipedia as well as the full German version of the Alpaca-GPT4 dataset (translated version).
Even though a lot of training has been done, this is still a tiny model and is highly limited to its small size. Expect many hallucinations and do not use this in a demanding production workflow.
# Links
- [Ollama](https://ollama.com/jace-ai/SmolLM2-German-Instruct)
- [Technical report paper](https://huggingface.co/mags0ft/SmolLM2-360m-German-Instruct/resolve/main/Enhancing%20Foreign%20Language%20Proficiency%20in%20SmolLM2-360M%20via%20Continued%20Pretraining%20and%20Instruction%20Fine-Tuning.pdf)
# Cite as
```bibtex
@misc{smollm2germaninstruct,
author = {Magnus Leonard Schlinsog},
title = {Enhancing Foreign Language Proficiency in SmolLM2-360M via Continued Pretraining and Instruction Fine-Tuning},
year = {2025},
url = {https://huggingface.co/mags0ft/SmolLM2-360m-German-Instruct},
}
```