OpenLLM-Ro/RoGemma-7b-Instruct-DPO

Model Card for Model ID

This model points/is identical to RoGemma-7b-Instruct-DPO-2024-10-09.

RoGemma is a family of pretrained and fine-tuned generative text models for Romanian. This is the repository for the human aligned instruct 7B model. Links to other models can be found at the bottom of this page.

Model Details

Model Description

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
• Language(s): Romanian
• License: cc-by-nc-4.0
• Finetuned from model: RoGemma-7b-Instruct-2024-10-09
• Trained using: RoHelpSteer

Model Sources

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

Intended Use

Intended Use Cases

RoGemma 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

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/RoGemma-7b-Instruct-DPO")
model = AutoModelForCausalLM.from_pretrained("OpenLLM-Ro/RoGemma-7b-Instruct-DPO")

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

Model	Average	ARC	MMLU	Winogrande	Hellaswag	GSM8k	TruthfulQA
gemma-1.1-7b-it	41.44	40.32	47.22	55.01	47.03	9.50	49.58
RoGemma-7b-Instruct-2024-06-28	53.41	52.44	54.44	69.36	61.96	31.06	51.23
RoGemma-7b-Instruct-2024-10-09	50.48	52.01	52.37	66.97	56.34	25.98	49.18
RoGemma-7b-Instruct-DPO-2024-10-09	48.27	46.66	54.45	63.73	49.33	34.98	40.45

Downstream tasks

	LaRoSeDa				WMT
	Few-shot		Finetuned		Few-shot		Finetuned
Model	Binary (Macro F1)	Multiclass (Macro F1)	Binary (Macro F1)	Multiclass (Macro F1)	EN-RO (Bleu)	RO-EN (Bleu)	EN-RO (Bleu)	RO-EN (Bleu)
gemma-1.1-7b-it	87.54	51.48	83.87	85.61	17.96	27.74	25.48	36.11
RoGemma-7b-Instruct-2024-06-28	97.86	65.70	98.43	87.17	27.91	23.08	27.99	39.51
RoGemma-7b-Instruct-2024-10-09	86.96	56.72	98.80	85.81	24.45	14.20	25.96	39.07
RoGemma-7b-Instruct-DPO-2024-10-09	96.45	63.23	-	-	20.73	7.87	-	-

	XQuAD				STS
	Few-shot		Finetuned		Few-shot		Finetuned
Model	(EM)	(F1)	(EM)	(F1)	(Spearman)	(Pearson)	(Spearman)	(Pearson)
gemma-1.1-7b-it	42.10	62.30	60.34	77.40	49.10	50.23	83.43	83.64
RoGemma-7b-Instruct-2024-06-28	17.75	28.11	52.02	68.43	73.96	75.16	86.45	86.31
RoGemma-7b-Instruct-2024-10-09	26.03	41.58	46.72	60.79	73.23	71.58	88.42	88.45
RoGemma-7b-Instruct-DPO-2024-10-09	19.14	38.10	-	-	69.38	69.34	-	-

MT-Bench

<

Model	Average	1st turn	2nd turn	Answers in Ro
gemma-1.1-7b-it	4.83	5.11	4.55	160/160
RoGemma-7b-Instruct-2024-06-28	5.26	5.92	4.60	160/160
RoGemma-7b-Instruct-2024-10-09	5.24	5.55	4.94	160/160
RoGemma-7b-Instruct-DPO-2024-10-09	5.47	5.92	5.03	160/160

RoCulturaBench

Model	Average	Answers in Ro
gemma-1.1-7b-it	3.38	100/100
RoGemma-7b-Instruct-2024-06-28	3.26	100/100
RoGemma-7b-Instruct-2024-10-09	3.51	100/100
RoGemma-7b-Instruct-DPO-2024-10-09	3.94	100/100

RoGemma Model Family

Model	Link
RoGemma-7b-Instruct-2024-06-28	link
RoGemma-7b-Instruct-2024-10-09	link
RoGemma-7b-Instruct-DPO-2024-10-09	link

Citation

@inproceedings{masala-etal-2024-vorbesti,
    title = "``Vorbe\c{s}ti Rom{\^a}ne\c{s}te?'' A Recipe to Train Powerful {R}omanian {LLM}s with {E}nglish Instructions",
    author = "Masala, Mihai and Ilie-Ablachim, Denis and Dima, Alexandru and Corlatescu, Dragos Georgian and Zavelca, Miruna-Andreea and Olaru, Ovio and Terian, Simina-Maria and Terian, Andrei and Leordeanu, Marius and Velicu, Horia and Popescu, Marius and Dascalu, Mihai and Rebedea, Traian",
    editor = "Al-Onaizan, Yaser and Bansal, Mohit and Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.681/",
    doi = "10.18653/v1/2024.findings-emnlp.681",
    pages = "11632--11647"
}