SmolLM2 is a family of compact language models available in three size: 135M, 360M, and 1.7B parameters. They are capable of solving a wide range of tasks while being lightweight enough to run on-device. More details in our paper: https://arxiv.org/abs/2502.02737

SmolLM2 demonstrates significant advances over its predecessor SmolLM1, particularly in instruction following, knowledge, reasoning. The 135M model was trained on 2 trillion tokens using a diverse dataset combination: FineWeb-Edu, DCLM, The Stack, along with new filtered datasets we curated and will release soon. We developed the instruct version through supervised fine-tuning (SFT) using a combination of public datasets and our own curated datasets. We then applied Direct Preference Optimization (DPO) using UltraFeedback.

The instruct model additionally supports tasks such as text rewriting, summarization and function calling (for the 1.7B) thanks to datasets developed by Argilla such as Synth-APIGen-v0.1. You can find the SFT dataset here: https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk and finetuning code at https://github.com/huggingface/alignment-handbook/tree/main/recipes/smollm2

How to use

pip install transformers

Running the model on CPU/GPU/multi GPU

Using full precision

python

# pip install transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "HuggingFaceTB/SmolLM2-135M"
device = "cuda" # for GPU usage or "cpu" for CPU usage
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# for multiple GPUs install accelerate and do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
inputs = tokenizer.encode("Gravity is", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

Using torch.bfloat16

python

# pip install accelerate
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
checkpoint = "HuggingFaceTB/SmolLM2-135M"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# for fp16 use `torch_dtype=torch.float16` instead
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", torch_dtype=torch.bfloat16)
inputs = tokenizer.encode("Gravity is", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

bash

>>> print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")
Memory footprint: 723.56 MB

Evaluation

In this section, we report the evaluation results of SmolLM2. All evaluations are zero-shot unless stated otherwise, and we use lighteval to run them.

Base pre-trained model

Metrics	SmolLM2-135M-8k	SmolLM-135M
HellaSwag	42.1	41.2
ARC (Average)	43.9	42.4
PIQA	68.4	68.4
MMLU (cloze)	31.5	30.2
CommonsenseQA	33.9	32.7
TriviaQA	4.1	4.3
Winogrande	51.3	51.3
OpenBookQA	34.6	34.0
GSM8K (5-shot)	1.4	1.0

Instruction model

Metric	SmolLM2-135M-Instruct	SmolLM-135M-Instruct
IFEval (Average prompt/inst)	29.9	17.2
MT-Bench	1.98	1.68
HellaSwag	40.9	38.9
ARC (Average)	37.3	33.9
PIQA	66.3	64.0
MMLU (cloze)	29.3	28.3
BBH (3-shot)	28.2	25.2
GSM8K (5-shot)	1.4	1.4

Limitations

SmolLM2 models primarily understand and generate content in English. They can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.

Training

Model

Architecture: Transformer decoder
Pretraining tokens: 2T
Precision: bfloat16

Hardware

GPUs: 64 H100

Software

Training Framework: nanotron

License

Apache 2.0

Citation