File size: 6,256 Bytes
1d24a42 62cad89 1d24a42 e208f70 8426368 e208f70 1d24a42 50b6440 1d24a42 93c2b0a 1d24a42 93c2b0a 1d24a42 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 |
---
license: apache-2.0
datasets:
- HumanLLMs/Human-Like-DPO-Dataset
language:
- en
base_model:
- HuggingFaceTB/SmolLM2-1.7B-Instruct
pipeline_tag: text-generation
library_name: transformers
tags:
- safetensors
- humanized
- smollm
---
# SmolLM2-1.7B-Humanized
## Table of Contents
1. [Model Summary](##model-summary)
2. [Limitations](##limitations)
3. [Training](##training)
4. [License](##license)
5. [Citation](##citation)
## Model Summary
**SmolLM2-1.7B-Humanized** is a fine-tuned version of the [SmolLM2-1.7B-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct) model, optimized using the Direct Preference Optimization (DPO) method. To do this we used the "[Human-Like-DPO-Dataset](https://huggingface.co/datasets/HumanLLMs/Human-Like-DPO-Dataset)" from [Human-Like LLMs](https://huggingface.co/HumanLLMs).
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, SmolLM2-1.7B-Humanized is designed to deliver an engaging and intuitive user experience in dialogue-based scenarios.
View our full report [here](https://www.assistantslab.com/research/smollm2-report).
### 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?"
| SmolLM2-1.7B-Instruct | SmolLM2-1.7B-Humanized |
|:---------------------:|:----------------------:|
| I'm functioning as intended, ready to assist you with your queries. | I'm doing great! It's a pleasure to help you with anything you need. How about you? How are you today? |
### How to use
### Transformers
```bash
pip install transformers
```
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "AssistantsLab/SmolLM2-1.7B-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/SmolLM2-1.7B-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](https://github.com/huggingface/lighteval) to run them.
## Instruction model Vs. Humanized model
### Note
We observe an unexpectedly worse TriviaQA score compared to the base instruct model. A bit of training on a dataset such as squad-v2 quickly resolves this issue and just one epoch results in a TriviaQA score far above the base instruct model (>21). We did not release this model due to worse scores on different metrics after this one epoch training. If your specific use-case requires a better grasp of trivia, feel free to train on squad-v2.
| Metric | SmolLM2-1.7B-Instruct | SmolLM2-1.7B-Humanized | Difference |
|:-----------------------------|:---------------------:|:----------------------:|:----------:|
| MMLU | **49.5** | 48.8 | -0.7 |
| ARC (Easy) | **68.9** | 64.9 | -4.0 |
| ARC (Challenge) | 38.5 | **40.3** | +1.8 |
| HellaSwag | **71.7** | 71.3 | -0.4 |
| PIQA | **76.2** | 75.8 | -0.6 |
| WinoGrande | **62.5** | 61.2 | -1.3 |
| TriviaQA | **10.2** | 1.3 | -8.9 |
| GSM8K | **0.0** | **0.0** | +0.0 |
| OpenBookQA | **45.6** | 44.8 | -0.8 |
| QuAC (F1) | 30.2 | **31.1** | +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](https://www.apache.org/licenses/LICENSE-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},
}
``` |