Model Card for GaMS3-12B-Instruct

GaMS3-12B-Instruct represents the next generation of the GaMS (Generative Model for Slovene) model. The model is based on Google's Gemma 3 family and continually pretrained on Slovene, English, and some portion of Croatian, Serbian, and Bosnian corpora. The supervised fine-tuning phase was done on a combination of Slovene and English datasets.

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 contributed to data collection and preparation, which enabled 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.

The model's development was supported by NVIDIA as a part of their Sovereign AI initiative. We are thankful for the access to NVIDIA DGX Cloud Lepton. We are also extremely grateful for all the support and help we received from a group of exceptional people at NVIDIA: Anna Louise Ollerenshaw, Meriem Bendris, Oleg Sudakov, Benedetta Delfino, Rita Fernandes Neves, Andrea Pilzer, Miguel Martinez, Noel Osagie, Adam Henryk Grzywaczewski and Aleks Polak.

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, Timotej Petrič, Dario Vajda and Marko Robnik-Šikonja.
• Languages: Slovene, English (primary), Croatian, Bosnian and Serbian (secondary). The model might also work for other languages supported by Gemma 3, even though it was not continually pretrained on them.
• Base model: google/gemma-3-12b-pt
• License: Gemma

Usage

Transformers library

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

from transformers import pipeline

model_id = "cjvt/GaMS3-12B-Instruct"

model = 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 = model(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 = model(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 (accelerate library required):

from transformers import pipeline

model_id = "cjvt/GaMS3-12B-Instruct"

model = 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 = model(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 = model(new_message, max_new_tokens=1024)
print("Model's response:", response[0]["generated_text"][-1]["content"])

vLLM library

As Gemma 3 architecture is supported in vLLM, this is also true for our model.

NOTE: We noticed degradation in performance when the Flash Infer attention backend is used. For optimal performance please use Flash Attention backend.

Example vLLM code:

from vllm import LLM, SamplingParams

model = LLM("cjvt/GaMS3-12B-Instruct")

sampling_params = SamplingParams(
    n=1,
    temperature=0.6,
    top_p=0.9,
    max_tokens=1024
)

messages = [[{"role": "user", "content": "Kateri je najpomembnejši dogodek v slovenski zgodovini?"}]]
response = model.chat(messages, sampling_params)
print("Model's response:", response[0].outputs[0].text)

Training

The training was performed in 3 CPT and 2 SFT stages.

CPT stages:

• Parallel alignment: the model was pretrained on parallel English and Slovene texts using context window of 65536 tokens;
• Base CPT: the model was pretrained on a combination of Slovene, English, Croatian, Bosnian and Serbian corpora with a context window of 65536 tokens;
• Long CPT: the model was pretrained on a combination of high quality Slovene, English, Croatian, Bosnian and Serbian corpora with a context window of 131072 tokens.

SFT stages:

• Base instruction-following SFT: the model was trained on a dataset consisting of various tasks (open/clsoed question answering, writing, math, code) and topics;
• Chat and safety tuning: the model was trained on a combination of chat-oriented examples and a small set of safety prompts.

Infrastructure

The model was trained on the following HPC infrastructure:

• EuroHPC supercomputer LEONARDO: We managed to scale the training across 128 nodes on LEONARDO's booster partition. We used approximately 150k GPU hours on LEONARDO for development of this model (including data preparation such as translation and web rewrite, and model training).
• Faculty's B200 node: With 8 B200 GPUs, our faculty's node represents a modern infrastructure for AI development. However, as we have only a single such node, the majority of the training was done elsewhere. In total, around 1000 GPU hours were used on the B200 node.
• NVIDIA DGX Cloud Lepton: A unified AI platform that connects developers to tens of thousands of GPUs from a global network of cloud providers. It addresses a critical need: accelerating AI developer productivity by providing access to GPU capacity and AI services across the NVIDIA compute ecosystem. It integrates seamlessly with the NVIDIA software stack, enabling developers to build, train, and deploy AI applications quickly and Scale. We spent approximately 40k Lepton GPU hours.

Software

• CPT: NVIDIA NeMo Framework 2.0 (container version 25.07);
• SFT: Transformers library in combination with DeepSpeed and TRL.

Training hyperparameters

In line with our commitment to transparency, open science, and the sharing of knowledge, we openly disclose all training hyperparameters used in developing this model. All training stages were performed with bfloat16 precision and Adam optimizer.

Stage	Model Parallelism	Data Parallelism	Batch Size	Micro Batch Size	LR Scheduler	Min LR	Max LR	Warmup Steps	Constant Steps	Epochs
Parallel alignment	TP 8	64	128	1	Cosine with warmup	5e-7	5e-6	150	200	1
Base CPT	TP 8	64	Rampup: 128 (961 steps) -> 192 (600 steps) -> 256	1	Cosine with warmup	5e-7	5e-6	1000	1000	1
Long CPT	TP 8	16	64	1	Constant with warmup	/	5e-6	500	/	1
Base instruction-following SFT	DeepSpeed ZeRO Stage 2	8	64	8	Cosine with warmup	1e-6	5e-6	1000	0	3 (checkpoint after epoch 2 was selected)
Chat and safety tuning	DeepSpeed ZeRO Stage 2	8	64	8	Cosine with warmup	1e-6	5e-6	1000	0	3 (checkpoint after epoch 2 was selected)

Data and benchmark information

Coming soon!

Usage and Limitations

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.