RoGemma2-9b-Instruct-DPO-2024-10-09

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

RoGemma2 is a family of pretrained and fine-tuned generative text models for Romanian. This is the repository for the human aligned instruct 9B model. Links to other models can be found at the bottom of this page.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

OpenLLM-Ro represents the first open-source effort to build a LLM specialized for Romanian. OpenLLM-Ro developed and publicly releases a collection of Romanian LLMs, both in the form of foundational model and instruct and chat variants.

• Developed by: OpenLLM-Ro

<!-- - **Funded by [optional]:** [More Information Needed] --> <!-- - **Shared by [optional]:** [More Information Needed] --> <!-- - **Model type:** [More Information Needed] -->

• Language(s): Romanian
• License: cc-by-nc-4.0
• Finetuned from model: RoGemma2-9b-Instruct-2024-10-09
• Trained using: RoHelpSteer

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: https://github.com/OpenLLM-Ro/LLaMA-Factory
• Paper: https://arxiv.org/abs/2406.18266

Intended Use

Intended Use Cases

RoGemma2 is intented for research use in Romanian. Base models can be adapted for a variety of natural language tasks while instruction and chat tuned models are intended for assistant-like chat.

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Use in any manner that violates the license, any applicable laws or regluations, use in languages other than Romanian.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2024-10-09")
model = AutoModelForCausalLM.from_pretrained("OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2024-10-09")

instruction = "Ce jocuri de societate pot juca cu prietenii mei?"
chat = [
        {"role": "user", "content": instruction},
        ]
prompt = tokenizer.apply_chat_template(chat, tokenize=False, system_message="")

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
outputs = model.generate(input_ids=inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0]))

Academic Benchmarks

<table> <tbody> <tr> <td><strong>Model</strong></td> <td><strong><center>Average</center></strong></td> <td><strong><center>ARC</center></strong></td> <td><strong><center>MMLU</center></strong></td> <td><strong><center>Winogrande</center></strong></td> <td><strong><center>Hellaswag</center></strong></td> <td><strong><center>GSM8k</center></strong></td> <td><strong><center>TruthfulQA</center></strong></td> </tr> <tr> <td>gemma-2-9b-it</td><td><center>56.22</center></td><td><center>50.33</center></td><td><center><strong>64.01</strong></center></td><td><center>64.88</center></td><td><center><strong>63.11</strong></center></td><td><center>41.95</center></td><td><center>53.03</center></td> </tr> <tr> <td>RoGemma2-9b-Instruct-2024-10-09</td><td><center>57.06</center></td><td><center><strong>56.20</strong></center></td><td><center>62.98</center></td><td><center><strong>71.00</strong></center></td><td><center>60.52</center></td><td><center>37.86</center></td><td><center><strong>53.77</strong></center></td> </tr> <tr> <td><em>RoGemma2-9b-Instruct-DPO-2024-10-09</em></td><td><center><em><strong>59.08</strong></em></center></td><td><center><em>54.10</em></center></td><td><center><em>63.41</em></center></td><td><center><em>70.02</em></center></td><td><center><em>59.35</em></center></td><td><center><em><strong>57.24</strong></em></center></td><td><center><em>50.39</em></center></td> </tr> </tbody> </table>

Downstream tasks

<table> <tbody> <tr> <td></td> <td colspan="4"><center><strong>LaRoSeDa</strong></center></td> <td colspan="4"><center><strong>WMT</strong></center></td> </tr> <tr> <td></td> <td colspan="2"><center><strong>Few-shot</strong></center></td> <td colspan="2"><center><strong>Finetuned</strong></center></td> <td colspan="2"><center><strong>Few-shot</strong></center></td> <td colspan="2"><center><strong>Finetuned</strong></center></td> </tr> <tr> <td><strong>Model</strong></td> <td><center><strong>Binary<br>(Macro F1)</strong></center></td> <td><center><strong>Multiclass<br>(Macro F1)</strong></center></td> <td><center><strong>Binary<br>(Macro F1)</strong></center></td> <td><center><strong>Multiclass<br>(Macro F1)</strong></center></td> <td><center><strong>EN-RO<br>(Bleu)</strong></center></td> <td><center><strong>RO-EN<br>(Bleu)</strong></center></td> <td><center><strong>EN-RO<br>(Bleu)</strong></center></td> <td><center><strong>RO-EN<br>(Bleu)</strong></center> </tr> <tr> <td>gemma-2-9b-it</td><td><center>90.82</center></td><td><center>52.51</center></td><td><center><strong>98.97</strong></center></td><td><center>86.02</center></td><td><center>19.97</center></td><td><center><strong>28.94</strong></center></td><td><center>27.94</center></td><td><center><strong>41.61</strong></center></td> </tr> <tr> <td>RoGemma2-9b-Instruct-2024-10-09</td><td><center>96.19</center></td><td><center>62.49</center></td><td><center>98.93</center></td><td><center><strong>88.33</strong></center></td><td><center>25.74</center></td><td><center>23.16</center></td><td><center><strong>28.43</strong></center></td><td><center>40.94</center></td> </tr> <tr> <td><em>RoGemma2-9b-Instruct-DPO-2024-10-09</em></td><td><center><em><strong>97.74</strong></em></center></td><td><center><em><strong>67.40</strong></em></center></td><td><center><em>-</em></center></td><td><center><em>-</em></center></td><td><center><em><strong>27.32</strong></em></center></td><td><center><em>15.96</em></center></td><td><center><em>-</em></center></td><td><center><em>-</em></center></td> </tr> </tbody> </table> <table> <tbody> <tr> <td></td> <td colspan="4"><center><strong>XQuAD</strong></center></td> <td colspan="4"><center><strong>STS</strong></center></td> </tr> <tr> <td></td> <td colspan="2"><center><strong>Few-shot</strong></center></td> <td colspan="2"><center><strong>Finetuned</strong></center></td> <td colspan="2"><center><strong>Few-shot</strong></center></td> <td colspan="2"><center><strong>Finetuned</strong></center></td> </tr> <tr> <td><strong>Model</strong></td> <td><center><strong>(EM)</strong></center></td> <td><center><strong>(F1)</strong></center></td> <td><center><strong>(EM)</strong></center></td> <td><center><strong>(F1)</strong></center></td> <td><center><strong>(Spearman)</strong></center></td> <td><center><strong>(Pearson)</strong></center></td> <td><center><strong>(Spearman)</strong></center></td> <td><center><strong>(Pearson)</strong></center></td> </tr> <tr> <td>gemma-2-9b-it</td><td><center>37.56</center></td><td><center>57.48</center></td><td><center><strong>71.09</strong></center></td><td><center><strong>84.78</strong></center></td><td><center>71.39</center></td><td><center>71.73</center></td><td><center>89.07</center></td><td><center>89.29</center></td> </tr> <tr> <td>RoGemma2-9b-Instruct-2024-10-09</td><td><center><strong>51.37</strong></center></td><td><center><strong>70.74</strong></center></td><td><center>50.00</center></td><td><center>64.10</center></td><td><center>77.15</center></td><td><center>77.10</center></td><td><center><strong>89.45</strong></center></td><td><center><strong>89.89</strong></center></td> </tr> <tr> <td><em>RoGemma2-9b-Instruct-DPO-2024-10-09</em></td><td><center><em>32.42</em></center></td><td><center><em>58.68</em></center></td><td><center><em>-</em></center></td><td><center><em>-</em></center></td><td><center><em><strong>80.82</strong></em></center></td><td><center><em><strong>81.50</strong></em></center></td><td><center><em>-</em></center></td><td><center><em>-</em></center></td> </tr> </tbody> </table>

MT-Bench

<table> <tbody> <tr> <td><strong>Model</strong></td> <td><strong><center>Average</center></strong></td> <td><strong><center>1st turn</center></strong></td> <td><strong><center>2nd turn</center></strong></td> <td><strong><center>Answers in Ro</center></strong></td> </tr> <tr> <td>gemma-2-9b-it</td><td><center><strong>7.50</strong></center></td><td><center><strong>7.91</strong></center></td><td><center><strong>7.09</strong></center></td><td><center>159/160</center></td> </tr> <tr> <td>RoGemma2-9b-Instruct-2024-10-09</td><td><center>6.08</center></td><td><center>6.78</center></td><td><center>5.39</center></td><td><center><strong>160/160</strong></center></td> </tr> <tr> <td><em>RoGemma2-9b-Instruct-DPO-2024-10-09</em></td><td><center><em>6.77</em></center></td><td><center><em>7.24</em></center></td><td><center><em>6.30</em></center></td><td><center><em><strong>160/160</strong></em></center></td> </tr> </tbody> </table>

RoCulturaBench

<table> <tbody> <tr> <td><strong>Model</strong></td> <td><strong><center>Average</center></strong></td> <td><strong><center>Answers in Ro</center></strong></td> </tr> <tr> <td>gemma-2-9b-it</td><td><center><strong>5.68</strong></center></td><td><center><strong>100/100</strong></center></td> </tr> <tr> <td>RoGemma2-9b-Instruct-2024-10-09</td><td><center>4.20</center></td><td><center><strong>100/100</strong></center></td> </tr> <tr> <td><em>RoGemma2-9b-Instruct-DPO-2024-10-09</em></td><td><center><em>4.83</em></center></td><td><center><em><strong>100/100</strong></em></center></td> </tr> </tbody> </table>

RoGemma2 Model Family

Citation

@misc{masala2024vorbecstiromanecsterecipetrain,
      title={"Vorbe\c{s}ti Rom\^ane\c{s}te?" A Recipe to Train Powerful Romanian LLMs with English Instructions}, 
      author={Mihai Masala and Denis C. Ilie-Ablachim and Alexandru Dima and Dragos Corlatescu and Miruna Zavelca and Ovio Olaru and Simina Terian-Dan and Andrei Terian-Dan and Marius Leordeanu and Horia Velicu and Marius Popescu and Mihai Dascalu and Traian Rebedea},
      year={2024},
      eprint={2406.18266},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.18266}, 
}

<!-- **APA:** [More Information Needed] -->