EX-Ai1-3.88B-Gemma3

EN — A LoRA fine-tuned Gemma 3 4B by EXAiStudio, optimized for Persian fluency, instruction-following, multilingual communication, intermediate Iranian cultural awareness, and reducing model hallucinations.

FA — نسخه فاین‌تیون‌شده Gemma 3 4B توسط EXAiStudio که برای بهبود روانی زبان فارسی، دستورپذیری، ارتباط چندزبانه، آشنایی متوسط با فرهنگ ایرانی و کاهش توهم‌های مدل بهینه‌سازی شده است.

ZH — 由 EXAiStudio 基于 Gemma 3 4B 进行 LoRA 微调的模型，重点提升波斯语流畅度、指令遵循、多语言交流能力、对伊朗文化的中级认知，并减少模型幻觉。

⚠️ Note on naming — The 3.88B in the model name refers to the effective parameter count of the Gemma 3 4B base model. Google markets it as "4B", but the actual parameter count is approximately 3.88B.

---

🇬🇧 English

📌 Model Details

Property	Value
Developed by	EXAiStudio
Model type	Causal Language Model (Decoder-only Transformer)
Base model	google/gemma-3-4b-it
Parameters	3.88B
Architecture	Gemma 3 (decoder-only Transformer, GQA, RoPE)
Context length	128,000 tokens (same as base model)
Fine-tuning method	LoRA
Language coverage	Multilingual, inherited from Gemma 3 pretraining
Fine-tuning focus	Persian fluency, instruction-following, Iranian culture, hallucination reduction
License	Gemma Terms of Use
Version	V1.0
Release date	October 22, 2026
Format	GGUF (Q4_K_M only)

🌐 Multilingual Coverage

EX-Ai1-3.88B-Gemma3 is derived from Gemma 3, a multilingual model family pretrained on data covering more than 140 languages.

The base model's multilingual capabilities may support communication, text generation, comprehension, and translation across a wide range of languages. The EXAiStudio fine-tuning specifically focuses on Persian fluency, instruction-following, Iranian cultural awareness, and reducing hallucinations.

Examples of languages within the broader multilingual landscape include:

• Major international languages: English, Spanish, French, German, Portuguese, Italian, Dutch, Russian, Chinese (Simplified and Traditional), Japanese, Korean, Arabic, Persian, Turkish, Hindi, Urdu, Bengali, Indonesian, Vietnamese, Thai, Swahili, and Filipino.
• European languages: Polish, Ukrainian, Romanian, Czech, Slovak, Hungarian, Bulgarian, Serbian, Croatian, Bosnian, Slovenian, Albanian, Greek, Swedish, Norwegian, Danish, Finnish, Icelandic, Estonian, Latvian, Lithuanian, Irish, Welsh, Scottish Gaelic, Catalan, Galician, Basque, Luxembourgish, Maltese, and Esperanto.
• South and Central Asian languages: Pashto, Kurdish, Tajik, Uzbek, Kazakh, Kyrgyz, Turkmen, Azerbaijani, Armenian, Georgian, Nepali, Sinhala, Punjabi, Gujarati, Marathi, Assamese, Odia, Tamil, Telugu, Kannada, Malayalam, and Mongolian.
• Southeast and East Asian languages: Burmese, Khmer, Lao, Javanese, Sundanese, Malay, Cebuano, Hmong, Tibetan, and Uyghur.
• African languages: Afrikaans, Amharic, Hausa, Yoruba, Igbo, Somali, Oromo, Kinyarwanda, Malagasy, Zulu, Xhosa, Shona, Sesotho, Setswana, Wolof, Bambara, and Lingala.
• Additional regional and historical languages: Latin, Yiddish, Māori, Hawaiian, Samoan, Tatar, Bashkir, Chechen, Ossetian, Breton, Occitan, Friulian, Sardinian, Romansh, and other regional languages.

This list is illustrative rather than exhaustive. It does not imply that every language listed has been individually tested or fine-tuned by EXAiStudio.

Important: Language coverage and language quality are not the same. Performance can vary considerably by language, dialect, writing system, task, and available training data. Persian is a primary focus of this fine-tuning, while equal performance across all languages is not guaranteed.

Description

EX-Ai1-3.88B-Gemma3 is a lightweight language model derived from Gemma 3 4B. It was fine-tuned with a LoRA adapter using a curated proprietary dataset of approximately 2,700 samples developed by NT Ai on commission for EXAiStudio.

The fine-tuning focused on:

• Strengthening Persian (Farsi) language quality and fluency.
• Improving instruction-following across structured tasks.
• Introducing intermediate Iranian cultural awareness, including customs, expressions, and cultural references.
• Reducing model hallucinations and improving the reliability of generated responses.
• Preserving and utilizing the multilingual capabilities inherited from the Gemma 3 base model.

The model is designed to provide a practical balance between response quality and computational requirements. With Q4_K_M quantization, it can be deployed on compatible consumer hardware, including mid-range GPUs and supported Apple Silicon systems.

Hallucination reduction is a training objective and an observed target of the fine-tuning process; it does not mean that the model is completely free of hallucinations or that every response is guaranteed to be correct.

🎯 Intended Use

✅ Direct Use

• Persian, English, and multilingual conversational assistants.
• Persian-language content generation and rewriting.
• Culturally aware Persian content generation.
• Structured instruction-following, including question answering and summarization.
• Translation and multilingual text assistance, subject to language-specific performance.
• General-purpose text generation and comprehension.
• Local and offline deployment through Ollama, llama.cpp, or LM Studio.
• Applications that benefit from reduced hallucination tendencies, with independent verification where accuracy is important.

❌ Out-of-Scope Use

• Highly advanced mathematics and symbolic computation without additional verification.
• Unverified medical, legal, financial, or psychological advice.
• High-stakes automated decision-making.
• Safety-critical systems.
• Treating generated responses as guaranteed facts.
• Assuming equal quality across all supported or inherited languages.

🏋️ Training Details

• Source: Proprietary dataset from NT Ai, commissioned by EXAiStudio.
• Dataset size: Approximately 2,700 curated samples.
• Fine-tuning method: LoRA.
• Focus areas: Persian fluency, instruction-following, Iranian cultural awareness, and hallucination reduction.
• Multilingual capability: Inherited from Gemma 3 pretraining.
• Maximum sequence length: Same as the base model.
• Hardware: 1 × NVIDIA Tesla T4 (16 GB VRAM).
• Training time: Approximately 1 hour and 45 minutes.
• Base model: Gemma 3 4B Instruct (full precision).

Fine-Tuning Objectives and Improvements

The fine-tuning process was designed to improve the base model in several areas:

1. Persian language quality: Improve fluency, phrasing, and natural language generation in Persian.
2. Instruction-following: Improve the model's ability to understand and follow user instructions, particularly structured tasks.
3. Iranian cultural awareness: Improve familiarity with Iranian customs, expressions, and cultural references.
4. Hallucination reduction: Reduce the tendency to generate fabricated or unsupported information and improve response reliability.
5. Multilingual usability: Retain the multilingual foundation of Gemma 3 while prioritizing the specific objectives of the EXAiStudio fine-tuning.

These improvements do not eliminate all model errors or guarantee factual accuracy. Independent evaluations are required to quantify the extent of improvement across languages and tasks.

Detailed hyperparameters, including LoRA rank, alpha, learning rate, epochs, and optimizer, are not disclosed at this time.

💻 How to Use

Ollama

ollama run hf.co/EXAiStudio/EX-Ai1-3.88B-Gemma3:Q4_K_M

llama.cpp

llama serve -hf EXAiStudio/EX-Ai1-3.88B-Gemma3:Q4_K_M

For additional deployment options, refer to the documentation for llama.cpp, Ollama, and Hugging Face.

---

Model Information and Limitations

• Model family: Gemma 3.
• Parameter count: Approximately 3.88 billion.
• Context window: Up to 128,000 tokens, inherited from the base model.
• Quantization: Q4_K_M.
• License: Subject to the applicable Gemma Terms of Use.
• Language performance: Varies by language and task.
• Hallucinations: Reduced hallucination tendencies are a fine-tuning objective, not a guarantee of factual correctness.

Developers should independently evaluate the model for their intended use cases before deployment, particularly where factual accuracy, multilingual reliability, or safety is important.

Acknowledgements

• Google DeepMind: For the Gemma 3 base model.
• NT Ai: For developing the proprietary dataset commissioned by EXAiStudio.
• EXAiStudio: For the fine-tuning, development, and release of EX-Ai1-3.88B-Gemma3.