Model Card for GaMS-9B-Instruct

GaMS-2B, GaMS-9B and GaMS-27B represent new improved and larger models of the GaMS (Generative Model for Slovene) familly. The models are based on Google's Gemma 2 familly and continually pretrained on Slovene, English and some portion of Croatian, Serbian and Bosnian corpora.

This is the SFT version of GaMS-9B model

Acknowledgment

The model was developed within the PoVeJMo research program (Adaptive Natural Language Processing with Large Language Models), particularly within the research project titled SloLLaMai -- Open-access computationally efficient models for Slovenian. The program is funded within the Recovery and Resilience Plan by the Slovenian Research and Innovation Agency (ARIS) and NextGenerationEU. The authors also acknowledge the financial support from the Slovenian Research and Innovation Agency (research core funding No. P6-0411 -- Language Resources and Technologies for Slovene).

We thank everyone who worked on data collection and preparation, enabling us to train our model. Special thanks go to Nikola Ljubešić, Taja Kuzman, Tjaša Arčon, Jaka Čibej, Simon Krek, Tomaž Erjavec, Iztok Kosem and Tomaž Savodnik.

Basic information

• Developed by: team of researchers at the University of Ljubljana, Faculty for Computer and Information Science. Team members: Domen Vreš, Iztok Lebar Bajec, Tjaša Arčon, Gašper Jelovčan and Marko Robnik-Šikonja.
• Languages: Slovene, English (primary), Croatian, Bosnian and Serbian (secondary). The model might also work for other languages supported by Gemma 2, even though it was not continually pretrained on them.
• Base model: cjvt/GaMS-9B
• License: Gemma

Usage

The model can be run through pipeline API using the following code:

from transformers import pipeline

model_id = "cjvt/GaMS-9B-Instruct"

pline = pipeline(
    "text-generation",
    model=model_id,
    device_map="cuda" # replace with "mps" to run on a Mac device
)

# Example of response generation
message = [{"role": "user", "content": "Kateri je najpomembnejši dogodek v slovenski zgodovini?"}]
response = pipeline(message, max_new_tokens=512)
print("Model's response:", response[0]["generated_text"][-1]["content"])

# Example of conversation chain
new_message = response[0]["generated_text"]
new_message.append({"role": "user", "content": "Lahko bolj podrobno opišeš ta dogodek?"})
response = pipeline(new_message, max_new_tokens=1024)
print("Model's response:", response[0]["generated_text"][-1]["content"])

For multi GPU inference set the device_map to auto:

from transformers import pipeline

model_id = "cjvt/GaMS-9B-Instruct"

pline = pipeline(
    "text-generation",
    model=model_id,
    device_map="auto"
)

# Example of response generation
message = [{"role": "user", "content": "Kateri je najpomembnejši dogodek v slovenski zgodovini?"}]
response = pipeline(message, max_new_tokens=512)
print("Model's response:", response[0]["generated_text"][-1]["content"])

# Example of conversation chain
new_message = response[0]["generated_text"]
new_message.append({"role": "user", "content": "Lahko bolj podrobno opišeš ta dogodek?"})
response = pipeline(new_message, max_new_tokens=1024)
print("Model's response:", response[0]["generated_text"][-1]["content"])

Data

CPT Data

Model was continually pre-trained in two stages. In the first stage, parallel English-Slovene (and Croatian in some cases) corpora was used to align the languages. In the second stage, the model was trained on separate English, Slovene, Croatian, Bosnian and Serbian corpora.

Parallel alignment corpora

Corpus	Alignment level	# Tokens	Percentage
KAS Abstracts	Document level	31 M	1.65 %
DGT	Separate documents	697 M	36.56 %
MaCoCu Parallel	Separate documents	430 M	22.53 %
CC-News	Paragraph level	749 M	39.25 %
Total		1.91 B

Explanation of each alignment level:

• Document level: Parallel documents were concatenated into a single document
• Separate documents: Parallel documents were not explicitly aligned
• Paragraph level: Paragraphs of parallel documents were concatenated (the first paragraph of Slovene/English document was followed by the first paragraph in the other language, which was then followed by the second paragraph in the first language and so on)

Second stage corpora

Corpus	Language	# Tokens	Percentage
KAS	Slovene	2.77 B	20.34 %
MetaFida*	Slovene	4.66 B	34.18 %
Wikipedia-En (Date: January 23rd 2025)	English	5.45 B	39.99 %
Wikipedia-Sl (Date: January 1st 2025)	Slovene	0.16 B	1.19 %
Wikipedia-Hr (Date: January 1st 2025)	Croatian	0.15 B	1.13 %
Wikipedia-Bs (Date: January 1st 2025)	Bosnian	0.07 B	0.50 %
Wikipedia-Sr-Latin*	Serbian	0.36 B	2.68 %
Total		13.62 B

Remarks:

• The following corpora was excluded from MetaFida: dgt15_sl, classlawiki_sl, tweet_sl, janes_tweet, janes_forum, janes_news
• Serbian Wikipedia was converted from Cyrillic to Latin

SFT Data

Our training data for SFT consisted out of approximately 25.000 training and 1500 validation examples. The dataset was a mixture of following datasets:

• GaMS-Instruct-GEN 1.0
• GaMS-Instruct-DH 1.0: 3000 randomly selected examples were chosen from this dataset
• GaMS-Instruct-MED 1.0: 3000 randomly selected examples were chosen from this dataset
• Parallel corpus EN-SL RSDO4 2.0: additional filtering was done on this corpus. First we ran FastText language identification using NeMo Curator and kept only the examples were source was detected as English and target as Slovene. Next, we ran COMET model to evaluate translations. We kept only the examples with COMET scores higher than 0.945 (approximatelly 8000 examples).
• Aya Dataset: only English and Serbian examples were taken from this dataset. Serbian examples were converted from Cyrillic to Latin.
• Math competitions: We took PDFs from Slovene national math competitions between years 2001 and 2010. We extracted text from PDFs using olmOCR and manually corrected the extracted text. This gave us around 150 solved math problems.

Training

The model was trained on the Booster partition of Leonardo HPC.

CPT

We continually pretrained the model using NVIDIA NeMo 2.0 framework. The model was trained in BF16-Mixed precision using tensor parallelism across 4 GPUs, sequence parallelism, and activation recomputation. The model was trained across 32 nodes, each containing 4 A100 64GB GPUs. The parallel alignment training took approximately 4 hours and second stage took approximately 40 hours.

The model was trained using a cosine learning rate scheduler with linear warmup and the following hyperparameters.

Parallel alignment:

• warmup steps: 150
• minimal learning rate: 5e-6
• maximal learning rate: 2e-5
• constant steps: 0
• batch size: 512 (4 million tokens)

Second stage:

• warmup steps: 500
• minimal learning rate: 5e-6
• maximal learning rate: 5e-5
• constant steps: 100
• batch size: 512 (4 million tokens)

SFT

For Supervised Fine-tuning we used Transformers library with DeepSpeed ZeRO-3. The model was trained in BF16 precision, using pipeline parallelism to split it across 4 GPUs. The model was trained on a single node with 4 A100 64 GB GPU. We used CPU offloading for the optimizer.

The model was tuned using a cosine learning rate scheduler with linear warmup and the following hyperparameters:

• number of epochs: training was done on 5 epochs, but the best performing model according to validation loss was obtained after the second epoch, so we kept that model
• batch size: 128
• warmup steps: 150
• minimal learning rate: 1e-7
• maximal learning rate: 5e-6
• constant steps: 0

Evaluation

The models were evaluated using Slovene SuperGLUE collection of classification tasks on SloBench. Instruct version of the model was also evaluated on translation from English to Slovene and from Slovene to English. Additionally, we evaluated our models on Slovenian-LLM-Eval.

Code for evaluation:

• SloBench tasks
• Slovenian-LLM-Eval

Slovenian-LLM-Eval results

Comparison between GaMS models, base Gemma 2 models and SlovenianGPT (open source model for Slovene based on Mistral 7B) is shown in the figure below. All models were evaluated in 0-shot scenario.

Slobench Results

GaMS 2B, 9B and 27B models were evaluated in 3-shot scenario, except for MultiRC and translation tasks, where 0-shot was used. GaMS-9B-Instruct was evaluated in 0-shot scenarion on all tasks. We used guided decoding to ensure the correct format of the responses.

Slovene SuperGLUE

Rank	Title	Average	BoolQ Accuracy	CB Accuracy	CB F1 Score	CB Average	COPA Accuracy	MultiRC EM	MultiRC F1a Score	MultiRC Average	RTE Accuracy	WSC Accuracy
1	GaMS-27B	0.7601	0.8333	0.6440	0.5864	0.6152	0.9540	0.3904	0.7504	0.5704	0.7931	0.7945
2	PrešernGPT 0.1	0.7568	0.8333	0.8520	0.5868	0.7194	0.9740	0.4985	0.8061	0.6523	0.8276	0.5342
3	Gemma 2 27B	0.7546	0.8333	0.6680	0.5972	0.6326	0.9140	0.4174	0.7295	0.5735	0.8276	0.7466
4	GaMS-9B	0.7309	0.7000	0.8400	0.7955	0.8178	0.9000	0.3243	0.6551	0.4897	0.7931	0.6849
5	GaMS-9B-Instruct	0.6997	0.8000	0.7960	0.7128	0.7544	0.8140	0.0721	0.6174	0.3447	0.7931	0.6918
6	Gemma 2 9B	0.6980	0.8333	0.8240	0.5683	0.6962	0.8700	0.2282	0.5310	0.3796	0.7241	0.6849
8	CroSloEngual BERT	0.6078	0.7333	0.7920	0.7437	0.7679	0.5720	0.0931	0.5241	0.3086	0.6552	0.6096
11	SlovenianGPT-Chat	0.5078	0.7333	0.3920	0.3829	0.3874	0.6840	0.2432	0.4944	0.3688	0.5172	0.3562
12	Gemma 2 2B	0.4876	0.6333	0.4520	0.2123	0.3321	0.5180	0.1471	0.4419	0.2945	0.5862	0.5616
13	GaMS-2B	0.4790	0.5667	0.6080	0.4880	0.5480	0.5240	0.0631	0.5234	0.2932	0.5517	0.3904
14	GaMS-1B	0.4604	0.5000	0.6200	0.4565	0.5382	0.4920	0.1351	0.2675	0.2013	0.4828	0.5479
15	GaMS-1B-Chat	0.4570	0.8000	0.4880	0.3023	0.3951	0.4840	0.1081	0.2428	0.1755	0.5172	0.3692

English to Slovene translation (first 11 models on the benchmark)

Rank	Title	BERT score	BLEU (avg)	METEOR (avg)	CHRF (avg)	BLEU (corpus)	CHRF (corpus)
1	DeepL Translator	0.8812	0.3153	0.5902	0.6205	0.3599	0.6205
2	gemini-1.5-pro	0.8791	0.3124	0.5895	0.6176	0.3569	0.6176
3	Sonnet 3.5	0.8789	0.3059	0.5783	0.6204	0.3442	0.6204
4	gpt-4o	0.8784	0.2958	0.5811	0.6138	0.3379	0.6138
5	EuroLLM-9B-Instruct	0.8741	0.2927	0.5792	0.6055	0.3273	0.6055
6	seamless-m4t-v2-large	0.8731	0.2780	0.5599	0.5947	0.3085	0.5947
7	GaMS-9B-Instruct	0.8713	0.2773	0.5616	0.5928	0.3209	0.5928
8	Zlatorog	0.8706	0.2834	0.5633	0.6014	0.2903	0.6014
9	RSDO-DS4-NMT 1.2.2	0.8705	0.2794	0.5634	0.5956	0.3226	0.5956
9	META LLAMA 3.1 405B	0.8705	0.2637	0.5497	0.5930	0.3063	0.5930
11	RSDO-DS4-NMT 1.2	0.8698	0.2781	0.5602	0.5970	0.3177	0.5970

Slovene to English translation (first 10 models on the benchmark)

Rank	Title	BERT score	BLEU (avg)	METEOR (avg)	CHRF (avg)	BLEU (corpus)	CHRF (corpus)
1	gpt-4o	0.9496	0.3161	0.6655	0.6297	0.3496	0.6297
2	gemini-1.5-pro	0.9489	0.3117	0.6560	0.6237	0.3502	0.6237
3	gpt-4o-mini	0.9466	0.2976	0.6493	0.6197	0.3328	0.6197
4	GaMS-9B-Instruct	0.9454	0.2821	0.6275	0.6018	0.3141	0.6018
5	ChatGPTv1	0.9449	0.2852	0.6415	0.6096	0.3171	0.6096
6	RSDO-DS4-NMT 1.2.4	0.9434	0.2839	0.6227	0.5967	0.3290	0.5967
7	RSDO-DS4-NMT 1.2.6	0.9433	0.2832	0.6207	0.5944	0.3295	0.5944
8	RSDO-DS4-NMT 1.2.2	0.9431	0.2785	0.6184	0.5933	0.3240	0.5933
8	RSDO-DS4-NMT 1.2	0.9431	0.2805	0.6201	0.5941	0.3231	0.5941
10	eTranslation SLEN	0.9414	0.2729	0.6175	0.5913	0.3119	0.5913

Usage and Limitations (taken from Gemma 2)

These models have certain limitations that users should be aware of.

Intended Usage

Open Large Language Models (LLMs) have a wide range of applications across various industries and domains. The following list of potential uses is not comprehensive. The purpose of this list is to provide contextual information about the possible use-cases that the model creators considered as part of model training and development.

• Content Creation and Communication
  • Text Generation: These models can be used to generate creative text formats such as poems, scripts, code, marketing copy, and email drafts.
  • Chatbots and Conversational AI: Power conversational interfaces for customer service, virtual assistants, or interactive applications.
  • Text Summarization: Generate concise summaries of a text corpus, research papers, or reports.
• Research and Education
  • Natural Language Processing (NLP) Research: These models can serve as a foundation for researchers to experiment with NLP techniques, develop algorithms, and contribute to the advancement of the field.
  • Language Learning Tools: Support interactive language learning experiences, aiding in grammar correction or providing writing practice.
  • Knowledge Exploration: Assist researchers in exploring large bodies of text by generating summaries or answering questions about specific topics.

Limitations

• Training Data
  • The quality and diversity of the training data significantly influence the model's capabilities. Biases or gaps in the training data can lead to limitations in the model's responses.
  • The scope of the training dataset determines the subject areas the model can handle effectively.
• Context and Task Complexity
  • LLMs are better at tasks that can be framed with clear prompts and instructions. Open-ended or highly complex tasks might be challenging.
  • A model's performance can be influenced by the amount of context provided (longer context generally leads to better outputs, up to a certain point).
• Language Ambiguity and Nuance
  • Natural language is inherently complex. LLMs might struggle to grasp subtle nuances, sarcasm, or figurative language.
• Factual Accuracy
  • LLMs generate responses based on information they learned from their training datasets, but they are not knowledge bases. They may generate incorrect or outdated factual statements.
• Common Sense
  • LLMs rely on statistical patterns in language. They might lack the ability to apply common sense reasoning in certain situations.

Ethical Considerations and Risks

The development of large language models (LLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following:

• Bias and Fairness
  • LLMs trained on large-scale, real-world text data can reflect socio-cultural biases embedded in the training material. These models underwent careful scrutiny, input data pre-processing described and posterior evaluations reported in this card.
• Misinformation and Misuse
  • LLMs can be misused to generate text that is false, misleading, or harmful.
  • Guidelines are provided for responsible use with the model, see the [Responsible Generative AI Toolkit][rai-toolkit].
• Transparency and Accountability:
  • This model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes.
  • A responsibly developed open model offers the opportunity to share innovation by making LLM technology accessible to developers and researchers across the AI ecosystem.

Risks identified and mitigations:

• Perpetuation of biases: It's encouraged to perform continuous monitoring (using evaluation metrics, human review) and the exploration of de-biasing techniques during model training, fine-tuning, and other use cases.
• Generation of harmful content: Mechanisms and guidelines for content safety are essential. Developers are encouraged to exercise caution and implement appropriate content safety safeguards based on their specific product policies and application use cases.
• Misuse for malicious purposes: Technical limitations and developer and end-user education can help mitigate against malicious applications of LLMs. Educational resources and reporting mechanisms for users to flag misuse are provided. Prohibited uses of Gemma models are outlined in the [Gemma Prohibited Use Policy][prohibited-use].
• Privacy violations: Models were trained on data filtered for removal of PII (Personally Identifiable Information). Developers are encouraged to adhere to privacy regulations with privacy-preserving techniques.