SaidzodaEng/Asad_Gemma3-27B_Adapter

Asad-Gemma3-27B-Tajik-Adapter

A LoRA adapter fine-tuned on 300M tokens of native Tajik educational content, enabling Gemma3-27B to understand Tajik syntax, culture, and pedagogy. Powers Ameena.tj — Central Asia’s first AI learning platform.

Model Details

Model Description

• Developed by: Saidzoda AI Research Lab (IT Park Tajikistan)
• Language(s): Tajik (primary), with cross-lingual support for Uzbek, Russian, English
• License: Apache 2.0
• Finetuned from: unsloth/gemma-3-27b-it-bnb-4bit (4-bit quantized)
• Model type: Instruction-tuned LLM adapter (LoRA)

Model Sources

• Repository: https://github.com/saidzoda/ameena-ai
• Demo: https://ameena.tj
• Organization: https://huggingface.co/SaidzodaEng

Uses

Direct Use

• AI tutoring in Tajik
• Homework explanation and generation
• Course creation for teachers

Downstream Use

Integrated into Ameena’s course generator, certification system, and offline PWA.

Out-of-Scope Use

• Not for medical, legal, or financial advice
• Not intended for high-stakes decision-making

Bias, Risks, and Limitations

• Reflects biases in the 300M-token Tajik corpus (textbooks, exams, literature)
• May struggle with rare dialects or highly informal speech
• Requires base model (gemma-3-27b-it-bnb-4bit) to run

How to Get Started

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "unsloth/gemma-3-27b-it-bnb-4bit"
adapter = "SaidzodaEng/Asad-Gemma3-27B-Tajik-Adapter"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    load_in_4bit=True,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter)

Training Details

Training Data

300M tokens of curated Tajik educational content: textbooks, exam questions, literary works, and technical manuals.

Training Procedure

Method: LoRA (rank=64, alpha=128)
Framework: Unsloth + PEFT
Precision: 4-bit (bnb)
Epochs: 0.3 (to avoid overfitting)
Environmental Impact

Hardware: Nvidia H200
Hours used: ~90
Cloud Provider: Runpod
Region: europe-west4
Carbon Emitted: ~0.9 kg CO₂eq (estimated via ML CO2 Impact Calculator)
Citation

APA:
Saidzoda AI Research Lab. (2025). Asad-Gemma3-27B-Tajik-Adapter. Hugging Face.

BibTeX:

@misc{saidzoda_asad_gemma3_2025,
  author = {Saidzoda AI Research Lab},
  title = {Asad-Gemma3-27B-Tajik-Adapter},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/SaidzodaEng/Asad-Gemma3-27B-Tajik-Adapter}}
}