OpenLLM-Ro/RoLlama3.1-8b-Instruct-DPO

Model Card for Model ID

Built with Meta Llama 3.1

This model points/is identical to RoLlama3.1-8b-Instruct-DPO-2024-10-09.

RoLlama3.1 is a family of pretrained and fine-tuned generative text models for Romanian. This is the repository for the human aligned instruct 8B 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: RoLlama3.1-8b-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

RoLlama3.1 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/RoLlama3.1-8b-Instruct-DPO")
model = AutoModelForCausalLM.from_pretrained("OpenLLM-Ro/RoLlama3.1-8b-Instruct-DPO")

instruction = "Ce jocuri de societate pot juca cu prietenii mei?"
chat = [
        {"role": "system", "content": "Ești un asistent folositor, respectuos și onest. Încearcă să ajuți cât mai mult prin informațiile oferite, excluzând răspunsuri toxice, rasiste, sexiste, periculoase și ilegale."},
        {"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
Llama-3.1-8B-Instruct	49.87	42.86	53.73	59.71	56.82	35.56	50.54
RoLlama3.1-8b-Instruct-2024-10-09	53.03	47.69	54.57	65.84	59.94	44.30	45.82
RoLlama3.1-8b-Instruct-DPO-2024-10-09	52.74	44.84	55.06	65.87	58.67	44.17	47.82

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)
Llama-3.1-8B-Instruct	95.74	59.49	98.57	82.41	19.01	27.77	29.02	39.80
RoLlama3.1-8b-Instruct-2024-10-09	94.56	60.10	95.12	87.53	21.88	23.99	28.27	40.44
RoLlama3.1-8b-Instruct-DPO-2024-10-09	96.10	55.37	-	-	21.29	21.86	-	-

	XQuAD				STS
	Few-shot		Finetuned		Few-shot		Finetuned
Model	(EM)	(F1)	(EM)	(F1)	(Spearman)	(Pearson)	(Spearman)	(Pearson)
Llama-3.1-8B-Instruct	44.96	64.45	69.50	84.31	72.11	71.64	84.59	84.96
RoLlama3.1-8b-Instruct-2024-10-09	13.59	23.56	49.41	62.93	75.89	76.00	86.86	87.05
RoLlama3.1-8b-Instruct-DPO-2024-10-09	21.58	36.54	-	-	78.01	77.98	-	-

MT-Bench

Model	Average	1st turn	2nd turn	Answers in Ro
Llama-3.1-8B-Instruct	5.69	5.85	5.53	160/160
RoLlama3.1-8b-Instruct-2024-10-09	5.42	5.95	4.89	160/160
RoLlama3.1-8b-Instruct-DPO-2024-10-09	6.21	6.74	5.69	160/160

RoCulturaBench

Model	Average	Answers in Ro
Llama-3.1-8B-Instruct	3.54	100/100
RoLlama3.1-8b-Instruct-2024-10-09	3.55	100/100
RoLlama3.1-8b-Instruct-DPO-2024-10-09	4.42	100/100

RoLlama3.1 Model Family

Model	Link
RoLlama3.1-8b-Instruct-2024-10-09	link
RoLlama3.1-8b-Instruct-DPO-2024-10-09	link

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}, 
}