Hugging Face has released its new SmolLM2 initiative, a line of tiny language models designed for fine-tuning on personal computers. The latest model, SmolLM2 1.7B, is trained on 11 tera-token datasets, significantly larger than the original SmolLM. This increase in data size allows for more accurate and informative results.
SmolLM2 1.7B comes with three variants: a base version and two instruct versions. The instruct versions are created using a similar recipe to train Zephyr, which utilizes supervised fine-tuning and distilled DPO (Direct Preference Optimization). This approach enables users to fine-tune their own models with direct preference optimization.
The new initiative also marks a significant milestone for Hugging Face, as they have fully released the pre-training data and recipe used in SmolLM2. This transparency ensures that the published evaluation results are accurate and reliable, making SmolLM2 an attractive option for researchers and developers looking to fine-tune their own models.
By providing a more accessible alternative to larger language models like Qwen2.5 and Llama 3.2, SmolLM2 offers a cost-effective solution for those seeking high-quality language processing capabilities on a budget.
Source: https://medium.com/@bnjmn_marie/smollm2-very-good-alternatives-to-qwen2-5-and-llama-3-2-463a200d2f3b