SmolLM2-135M-Humanized is a fine-tuned version of the SmolLM2-135M-Instruct model, optimized using the Direct Preference Optimization (DPO) method. To do this we used the "Human-Like-DPO-Dataset" from Human-Like LLMs. To not lose too much quality with this post-training, we also applied some extra training on the "openbmb/UltraFeedback" dataset.

Unlike traditional fine-tuning datasets that aim to improve specific benchmarks or metrics, the Human-Like-DPO-Dataset focuses on aligning the model's behavior with human preferences. This process enhances the model's ability to generate more natural, human-like responses, making it particularly well-suited for conversational applications.

By emphasizing response quality and relatability, SmolLM2-135M-Humanized is designed to deliver an engaging and intuitive user experience in dialogue-based scenarios.

View our full report here.

Model example response

To reiterate, the goal is to make the model more human and less 'robot like'. Given the system prompt "You are a helpful assistant that lives inside the users PC" and the user message "How are you?"

SmolLM2-135M-Instruct	SmolLM2-135M-Humanized
I'm doing great! I'm a helpful assistant that lives inside users PC. I'm here to help you with any questions or problems you might have.	I'm doing great! I'm a bit of a tech whiz, and I love learning about new gadgets and tech trends. What's on your mind? Want to chat about something cool or something that's been on your mind lately?

How to use

Transformers

pip install transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "AssistantsLab/SmolLM2-135M-humanized"

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 gravity?"}]
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:

bash

pip install trl
trl chat --model_name_or_path AssistantsLab/SmolLM2-135M-humanized --device cpu

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.

Instruction model Vs. Humanized model

Metric	SmolLM2-135M-Instruct	SmolLM2-135M-Humanized	Difference
MMLU	23.1	23.0	-0.1
ARC (Easy)	54.3	55.0	+0.7
ARC (Challenge)	26.1	25.5	-0.6
HellaSwag	43.0	42.4	-0.6
PIQA	67.2	67.0	-0.2
WinoGrande	52.5	52.1	-0.4
TriviaQA	0.3	0.2	-0.1
GSM8K	0.2	0.8	+0.6
OpenBookQA	32.6	33.0	+0.4
QuAC (F1)	14.1	13.2	-0.9

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.

Humanized models display a bigger preference for confident hallucinating in some limited testing. Please keep this in mind in any potential applications.

License

Apache 2.0

Citation

SmolLM2:

bash

@misc{allal2024SmolLM2,
      title={SmolLM2 - with great data, comes great performance}, 
      author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Gabriel Martín Blázquez and Lewis Tunstall and Agustín Piqueres and Andres Marafioti and Cyril Zakka and Leandro von Werra and Thomas Wolf},
      year={2024},
}

Human-Like-DPO-Dataset:

bash

@misc{çalık2025enhancinghumanlikeresponseslarge,
      title={Enhancing Human-Like Responses in Large Language Models}, 
      author={Ethem Yağız Çalık and Talha Rüzgar Akkuş},
      year={2025},
      eprint={2501.05032},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.05032}, 
}

UltraFeedback dataset:

bash

@misc{cui2023ultrafeedback,
      title={UltraFeedback: Boosting Language Models with High-quality Feedback}, 
      author={Ganqu Cui and Lifan Yuan and Ning Ding and Guanming Yao and Wei Zhu and Yuan Ni and Guotong Xie and Zhiyuan Liu and Maosong Sun},
      year={2023},
      eprint={2310.01377},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

AssistantsLab/SmolLM2-135M-humanized

SmolLM2-135M-Humanized

Table of Contents

Model Summary

Model example response

How to use

Transformers

Chat in TRL

Evaluation

Instruction model Vs. Humanized model

Limitations

License

Citation

AssistantsLab/
SmolLM2-135M-humanized