--- license: cc-by-nc-4.0 language: - ro base_model: - google/gemma-2-9b-it datasets: - OpenLLM-Ro/ro_sft_alpaca - OpenLLM-Ro/ro_sft_alpaca_gpt4 - OpenLLM-Ro/ro_sft_dolly - OpenLLM-Ro/ro_sft_selfinstruct_gpt4 - OpenLLM-Ro/ro_sft_norobots - OpenLLM-Ro/ro_sft_orca - OpenLLM-Ro/ro_sft_camel - OpenLLM-Ro/ro_sft_oasst - OpenLLM-Ro/ro_sft_ultrachat - OpenLLM-Ro/ro_sft_magpie_mt - OpenLLM-Ro/ro_sft_magpie_reasoning model-index: - name: OpenLLM-Ro/RoGemma2-9b-Instruct-2025-04-23 results: - task: type: text-generation dataset: name: RoMT-Bench type: RoMT-Bench metrics: - name: Score type: Score value: 6.78 - task: type: text-generation dataset: name: RoCulturaBench type: RoCulturaBench metrics: - name: Score type: Score value: 4.89 - task: type: text-generation dataset: name: Romanian_Academic_Benchmarks type: Romanian_Academic_Benchmarks metrics: - name: Average accuracy type: accuracy value: 54.39 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_arc_challenge type: OpenLLM-Ro/ro_arc_challenge metrics: - name: Average accuracy type: accuracy value: 50.24 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_mmlu type: OpenLLM-Ro/ro_mmlu metrics: - name: Average accuracy type: accuracy value: 62.00 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_winogrande type: OpenLLM-Ro/ro_winogrande metrics: - name: Average accuracy type: accuracy value: 70.38 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_hellaswag type: OpenLLM-Ro/ro_hellaswag metrics: - name: Average accuracy type: accuracy value: 52.25 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_gsm8k type: OpenLLM-Ro/ro_gsm8k metrics: - name: Average accuracy type: accuracy value: 40.51 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_truthfulqa type: OpenLLM-Ro/ro_truthfulqa metrics: - name: Average accuracy type: accuracy value: 50.97 - task: type: text-generation dataset: name: LaRoSeDa_binary type: LaRoSeDa_binary metrics: - name: Average macro-f1 type: macro-f1 value: 84.23 - task: type: text-generation dataset: name: LaRoSeDa_multiclass type: LaRoSeDa_multiclass metrics: - name: Average macro-f1 type: macro-f1 value: 60.14 - task: type: text-generation dataset: name: WMT_EN-RO type: WMT_EN-RO metrics: - name: Average bleu type: bleu value: 17.78 - task: type: text-generation dataset: name: WMT_RO-EN type: WMT_RO-EN metrics: - name: Average bleu type: bleu value: 18.24 - task: type: text-generation dataset: name: XQuAD type: XQuAD metrics: - name: Average exact_match type: exact_match value: 49.22 - task: type: text-generation dataset: name: XQuAD type: XQuAD metrics: - name: Average f1 type: f1 value: 66.33 - task: type: text-generation dataset: name: STS type: STS metrics: - name: Average spearman type: spearman value: 70.17 - task: type: text-generation dataset: name: STS type: STS metrics: - name: Average pearson type: pearson value: 70.80 - task: type: text-generation dataset: name: RoMT-Bench type: RoMT-Bench metrics: - name: First turn type: Score value: 7.00 - name: Second turn type: Score value: 6.55 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_arc_challenge type: OpenLLM-Ro/ro_arc_challenge metrics: - name: 0-shot type: accuracy value: 47.47 - name: 1-shot type: accuracy value: 50.56 - name: 3-shot type: accuracy value: 50.73 - name: 5-shot type: accuracy value: 50.39 - name: 10-shot type: accuracy value: 50.99 - name: 25-shot type: accuracy value: 51.33 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_mmlu type: OpenLLM-Ro/ro_mmlu metrics: - name: 0-shot type: accuracy value: 58.73 - name: 1-shot type: accuracy value: 60.12 - name: 3-shot type: accuracy value: 64.93 - name: 5-shot type: accuracy value: 64.21 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_winogrande type: OpenLLM-Ro/ro_winogrande metrics: - name: 0-shot type: accuracy value: 66.06 - name: 1-shot type: accuracy value: 70.40 - name: 3-shot type: accuracy value: 72.30 - name: 5-shot type: accuracy value: 72.77 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_hellaswag type: OpenLLM-Ro/ro_hellaswag metrics: - name: 0-shot type: accuracy value: 56.30 - name: 1-shot type: accuracy value: 58.29 - name: 3-shot type: accuracy value: 50.88 - name: 5-shot type: accuracy value: 44.38 - name: 10-shot type: accuracy value: 51.41 - task: type: text-generation dataset: name: OpenLLM-Ro/ro_gsm8k type: OpenLLM-Ro/ro_gsm8k metrics: - name: 1-shot type: accuracy value: 27.29 - name: 3-shot type: accuracy value: 39.04 - name: 5-shot type: accuracy value: 55.19 - task: type: text-generation dataset: name: LaRoSeDa_binary type: LaRoSeDa_binary metrics: - name: 0-shot type: macro-f1 value: 59.19 - name: 1-shot type: macro-f1 value: 94.22 - name: 3-shot type: macro-f1 value: 93.24 - name: 5-shot type: macro-f1 value: 90.27 - task: type: text-generation dataset: name: LaRoSeDa_multiclass type: LaRoSeDa_multiclass metrics: - name: 0-shot type: macro-f1 value: 32.52 - name: 1-shot type: macro-f1 value: 68.64 - name: 3-shot type: macro-f1 value: 70.14 - name: 5-shot type: macro-f1 value: 69.26 - task: type: text-generation dataset: name: WMT_EN-RO type: WMT_EN-RO metrics: - name: 0-shot type: bleu value: 1.96 - name: 1-shot type: bleu value: 27.30 - name: 3-shot type: bleu value: 28.31 - name: 5-shot type: bleu value: 13.56 - task: type: text-generation dataset: name: WMT_RO-EN type: WMT_RO-EN metrics: - name: 0-shot type: bleu value: 0.66 - name: 1-shot type: bleu value: 26.76 - name: 3-shot type: bleu value: 31.88 - name: 5-shot type: bleu value: 13.66 - task: type: text-generation dataset: name: XQuAD_EM type: XQuAD_EM metrics: - name: 0-shot type: exact_match value: 49.92 - name: 1-shot type: exact_match value: 47.98 - name: 3-shot type: exact_match value: 45.71 - name: 5-shot type: exact_match value: 53.28 - task: type: text-generation dataset: name: XQuAD_F1 type: XQuAD_F1 metrics: - name: 0-shot type: f1 value: 67.52 - name: 1-shot type: f1 value: 63.97 - name: 3-shot type: f1 value: 62.39 - name: 5-shot type: f1 value: 71.43 - task: type: text-generation dataset: name: STS_Spearman type: STS_Spearman metrics: - name: 1-shot type: spearman value: 82.53 - name: 3-shot type: spearman value: 65.73 - name: 5-shot type: spearman value: 62.25 - task: type: text-generation dataset: name: STS_Pearson type: STS_Pearson metrics: - name: 1-shot type: pearson value: 82.89 - name: 3-shot type: pearson value: 66.26 - name: 5-shot type: pearson value: 63.25 --- # Model Card for Model ID This model points/is identical to [RoGemma2-9b-Instruct-2025-04-23](https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct-2025-04-23). RoGemma2 is a family of pretrained and fine-tuned generative text models for Romanian. This is the repository for the **instruct 9B 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:** [gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it) - **Trained using:** [RoAlpaca](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_alpaca), [RoAlpacaGPT4](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_alpaca_gpt4), [RoDolly](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_dolly), [RoSelfInstruct](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_selfinstruct_gpt4), [RoNoRobots](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_norobots), [RoOrca](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_orca), [RoCamel](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_camel), [RoOpenAssistant](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_oasst), [RoUltraChat](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_ultrachat), [RoMagpiePro](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_magpie_mt), [RoMagpieReasoning](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_magpie_reasoning) ### Model Sources - **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 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. ```python from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("OpenLLM-Ro/RoGemma2-9b-Instruct") model = AutoModelForCausalLM.from_pretrained("OpenLLM-Ro/RoGemma2-9b-Instruct") 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-2-9b-it
56.22
50.33
64.01
64.88
63.11
41.95
53.03
RoGemma2-9b-Instruct-2024-10-09
57.06
56.20
62.98
71.00
60.52
37.86
53.77
RoGemma2-9b-Instruct-2025-04-23
54.39
50.24
62.00
70.38
52.25
40.51
50.97
RoGemma2-9b-Instruct-DPO-2024-10-09
59.08
54.10
63.41
70.02
59.35
57.24
50.39
RoGemma2-9b-Instruct-DPO-2025-04-23
59.79
55.66
64.00
73.16
64.26
37.80
63.86
## 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-2-9b-it
90.82
52.51
98.97
86.02
19.97
28.94
27.94
41.61
RoGemma2-9b-Instruct-2024-10-09
96.19
62.49
98.93
88.33
25.74
23.16
28.43
40.94
RoGemma2-9b-Instruct-2025-04-23
84.23
60.14
-
-
17.78
18.24
-
-
RoGemma2-9b-Instruct-DPO-2024-10-09
97.74
67.40
-
-
27.32
15.96
-
-
RoGemma2-9b-Instruct-DPO-2025-04-23
82.84
65.95
-
-
28.16
19.34
-
-
XQuAD
STS
Few-shot
Finetuned
Few-shot
Finetuned
Model
(EM)
(F1)
(EM)
(F1)
(Spearman)
(Pearson)
(Spearman)
(Pearson)
gemma-2-9b-it
37.56
57.48
71.09
84.78
71.39
71.73
89.07
89.29
RoGemma2-9b-Instruct-2024-10-09
51.37
70.74
50.00
64.10
77.15
77.10
89.45
89.89
RoGemma2-9b-Instruct-2025-04-23
49.22
66.33
-
-
70.17
70.80
-
-
RoGemma2-9b-Instruct-DPO-2024-10-09
32.42
58.68
-
-
80.82
81.50
-
-
RoGemma2-9b-Instruct-DPO-2025-04-23
30.82
48.53
-
-
73.24
73.13
-
-
## MT-Bench
Model
Average
1st turn
2nd turn
Answers in Ro
gemma-2-9b-it
7.50
7.91
7.09
159/160
RoGemma2-9b-Instruct-2024-10-09
6.08
6.78
5.39
160/160
RoGemma2-9b-Instruct-2025-04-23
6.78
7.00
6.55
160/160
RoGemma2-9b-Instruct-DPO-2024-10-09
6.77
7.24
6.30
160/160
RoGemma2-9b-Instruct-DPO-2025-04-23
7.26
7.65
6.86
160/160
## RoCulturaBench
Model
Average
Answers in Ro
gemma-2-9b-it
5.68
100/100
RoGemma2-9b-Instruct-2024-10-09
4.20
100/100
RoGemma2-9b-Instruct-2025-04-23
4.89
100/100
RoGemma2-9b-Instruct-DPO-2024-10-09
4.83
100/100
RoGemma2-9b-Instruct-DPO-2025-04-23
5.36
100/100
## RoGemma2 Model Family | Model | Link | |--------------------|:--------:| |RoGemma2-9b-Instruct-2024-10-09| [link](https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct-2024-10-09) | |*RoGemma2-9b-Instruct-2025-04-23*| [link](https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct-2024-10-09) | |RoGemma2-9b-Instruct-DPO-2024-10-09| [link](https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2024-10-09) | |RoGemma2-9b-Instruct-DPO-2025-04-23| [link](https://huggingface.co/OpenLLM-Ro/RoGemma2-9b-Instruct-DPO-2024-10-09) | ## 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}, } ```