unsloth/SmolLM-135M-Instruct-bnb-4bit

Finetune Llama 3.1, Gemma 2, Mistral 2-5x faster with 70% less memory via Unsloth!

We have a free Google Colab Tesla T4 notebook for Llama 3.1 (8B) here - also works for SmolLM!: https://colab.research.google.com/drive/1Ys44kVvmeZtnICzWz0xgpRnrIOjZAuxp?usp=sharing

✨ Finetune for Free

All notebooks are beginner friendly! Add your dataset, click "Run All", and you'll get a 2x faster finetuned model which can be exported to GGUF, vLLM or uploaded to Hugging Face.

Unsloth supports	Free Notebooks	Performance	Memory use
Llama-3.1 8b	▶️ Start on Colab	2.4x faster	58% less
Phi-3.5 (mini)	▶️ Start on Colab	2x faster	50% less
Gemma-2 9b	▶️ Start on Colab	2.4x faster	58% less
Mistral 7b	▶️ Start on Colab	2.2x faster	62% less
TinyLlama	▶️ Start on Colab	3.9x faster	74% less
DPO - Zephyr	▶️ Start on Colab	1.9x faster	19% less

• This conversational notebook is useful for ShareGPT ChatML / Vicuna templates.
• This text completion notebook is for raw text. This DPO notebook replicates Zephyr.
• * Kaggle has 2x T4s, but we use 1. Due to overhead, 1x T4 is 5x faster.

SmolLM-1.7B-Instruct

Model Summary

SmolLM is a series of small language models available in three sizes: 135M, 360M, and 1.7B parameters.

These models are pre-trained on SmolLM-Corpus, a curated collection of high-quality educational and synthetic data designed for training LLMs. For further details, we refer to our blogpost.

To build SmolLM-Instruct, we finetuned the base models on publicly available datasets.

Changelog

Release	Description
v0.1	Initial release of SmolLM-Instruct. We finetune on the permissive subset of the WebInstructSub dataset, combined with StarCoder2-Self-OSS-Instruct. Then, we perform DPO (Direct Preference Optimization) for one epoch on HelpSteer for the 135M and 1.7B models, and argilla/dpo-mix-7k for the 360M model.
v0.2	We changed the finetuning mix to datasets more suitable for smol models. We train on a new dataset of 2k simple everyday conversations we generated by llama3.1-70B everyday-conversations-llama3.1-2k, Magpie-Pro-300K-Filtered, StarCoder2-Self-OSS-Instruct, and a small subset of OpenHermes-2.5

v0.2 models are better at staying on topic and responding appropriately to standard prompts, such as greetings and questions about their role as AI assistants. SmolLM-360M-Instruct (v0.2) has a 63.3% win rate over SmolLM-360M-Instruct (v0.1) on AlpacaEval. You can find the details here.

You can load v0.1 checkpoint by specifying revision="v0.1" in the transformers code:

model = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM-1.7B-Instruct", revision="v0.1")

Usage

Local Applications

⚡ For local applications, you can find optimized implementations of the model in MLC, GGUF and Transformers.js formats, in addition to fast in-browser demos in this collection: https://huggingface.co/collections/HuggingFaceTB/local-smollms-66c0f3b2a15b4eed7fb198d0

We noticed that 4bit quantization degrades the quality of the 135M and 360M, so we use q016 for MLC and ONNX/Transformers.js checkpoints for the WebGPU demos. We also suggest using temperature 0.2 and top-p 0.9.

Transformers

pip install transformers

# pip install transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "HuggingFaceTB/SmolLM-1.7B-Instruct"

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)

messages = [{"role": "user", "content": "What is the capital of France."}]
input_text=tokenizer.apply_chat_template(messages, tokenize=False)
print(input_text)
inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
outputs = model.generate(inputs, max_new_tokens=50, temperature=0.2, top_p=0.9, do_sample=True)
print(tokenizer.decode(outputs[0]))

Chat in TRL

You can also use the TRL CLI to chat with the model from the terminal:

pip install trl
trl chat --model_name_or_path HuggingFaceTB/SmolLM-1.7B-Instruct --device cpu

Limitations

Additionally, the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data, we invite users to leverage them as assistive tools rather than definitive sources of information. We find that they can handle general knowledge questions, creative writing and basic Python programming. But they are English only and may have difficulty with arithmetics, editing tasks and complex reasoning. For more details about the models' capabilities, please refer to our blog post.

Training parameters

We train the models using the alignment-handbook with the datasets mentioned in the changelog, using these parameters v0.2 (most of them are from Zephyr Gemma recipe):

• 1 epoch
• lr 1e-3
• cosine schedule
• warmup ratio 0.1
• global batch size 262k tokens

You can find the training recipe here: https://github.com/huggingface/alignment-handbook/tree/smollm/recipes/smollm

Citation

@misc{allal2024SmolLM,
      title={SmolLM - blazingly fast and remarkably powerful}, 
      author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Leandro von Werra and Thomas Wolf},
      year={2024},
}