This repository contains the fine-tuned LoRA adapters for Gemma-3-4b-it, trained specifically to generate accurate, context-aware Lithuanian descriptions and captions for visual data. This project was developed as part of the LT-VLM pilot initiative in collaboration with Vilnius University and databl.ai.

Model Details

Developed by: Vilnius University Student Project Team
Model type: Vision-Language Model (VLM) LoRA Adapter
Language: Lithuanian (lt)
Finetuned from model: unsloth/gemma-3-4b-it-bnb-4bit

Project Context

Curation Platform: databl.ai LT-VLM
Base Dataset Repository: GoAGI-AI/lt-vlm

Model Description

This model is a fine-tuned LoRA adapter based on Gemma-3-4b-it that specializes in identifying and captioning Lithuanian churches and religious architecture. It is designed to look at photos of local churches, chapels, and cathedrals, and then generate highly accurate, grammatically correct captions in native Lithuanian. By focusing on this targeted architectural domain, it ensures proper recognition of specific historical heritage sites, architectural styles (such as Lithuanian Gothic or Baroque), and correct native proper nouns with their appropriate diacritics (ė, š, ų, ž).

Developed by: Vilnius University Student Project Team "Cento per cento"
Model type: Vision-Language Model (VLM) LoRA Adapter
Language: Lithuanian (lt)
License: MIT
Finetuned from model: unsloth/gemma-3-4b-it-bnb-4bit

Model Sources

Repository: https://huggingface.co/AugusteBu/Cento_per_cento

Uses

Direct Use

This model is intended to receive an image of a Lithuanian church exterior and output a structurally sound description in Lithuanian. It is optimized to recognize specific regional landmarks (e.g., Šv. Onos bažnyčia, Vilniaus arkikatedra) and describe architectural elements like belfries, altars, or facades much more accurately than generic factory-default models.

Bias, Risks, and Limitations

The model was trained on a limited set of ~100+ images. Its performance may vary on poorly lit, highly atypical, or heavily obstructed church images.
It may occasionally misidentify a church or use generic descriptions when faced with very rare architectural styles.
Like all language models, it can produce factually incorrect statements if the image is ambiguous.
No systematic bias evaluation has been conducted regarding denomination, age, or geographic distribution of churches.

Recommendations

Users should validate outputs when used in critical applications. The model is best used as an assistive tool alongside human verification.

Training Details

Training Data

The training pipeline utilized a custom-curated subset of 100+ images of Lithuanian scenes. The data lifecycle began with crowdsourced submissions via Discord, followed by quality triage, auto-captioning evaluation, and final verification by native Lithuanian speakers within the databl.ai data environment.

Training Progress & Loss

(Tip: You can directly upload or drag-and-drop a screenshot of your notebook's training loss graph right here via the Hugging Face Web Editor to showcase your optimization metrics!)

Training Procedure

The model was fine-tuned using the Unsloth library, leveraging QLoRA (Quantized Low-Rank Adaptation) to efficiently attach lightweight adapters to the quantized 4-bit base vision model.

Preprocessing

Images were automatically downscaled, normalized, and patched using the native image processor for gemma-3-4b-it. Textual descriptions were tokenized and structured alongside the image inputs using the model's standardized chat template configuration.

Training Hyperparameters

Training regime: bf16 mixed precision (or fp16 depending on the notebook GPU) utilizing 4-bit base model quantization (bitsandbytes).
Optimizer: adamw_8bit
LoRA Rank (r): 16
LoRA Alpha (alpha): 16
Target Modules: Vision and language attention projection layers.

Speeds, Sizes, Times

Dataset Size: 100+ curated Lithuanian scene images and native captions.
Final Adapter Weights Size: ~116 MB (adapter_model.safetensors).
Training Duration: ~5 to 15 minutes depending on the active notebook GPU (e.g., NVIDIA T4 or A100).

Evaluation

Testing Data, Factors & Metrics

Testing Data

Evaluation was performed qualitatively using a held-out set of test images featuring distinct Lithuanian categories (e.g., historical landmarks, local street signage, and traditional food items) that were not seen by the model during the training loop. This directly supports the project's requirement to demonstrate inference improvements.

Factors

The model evaluation focused on two primary factors:

Linguistic Precision: The correct implementation of Lithuanian diacritics (ė, š, ų, ž), grammatical cases, and natural sentence structures.
Cultural/Contextual Accuracy: The ability to correctly train and try to identify lithuanian churches(e.g., identifying Vilniaus Šv. Kotrynos bažnyčia)

Metrics

Given the specific nature of this baseline pilot, progress was evaluated using:

Qualitative Comparative Inference: A direct side-by-side assessment of generated text outputs comparing the base model baseline ("Before Fine-Tuning") against the newly trained adapter ("After Fine-Tuning").

Results

Summary

The model wsa trained on chucrces and managed to detect very few churches correctly. While it improved being able to detect that they are Lithuanian churches. It could not correcly identify most of them.

Technical Specifications

Software

Training framework: Unsloth
Libraries: transformers 4.46+, peft 0.13+, torch 2.4+, unsloth 2024.12+
Environment: Google Colab notebook with Python 3.10

BibTeX:

@misc{ltvlm_churches_2026, author = {Vilnius University Student Project Team "Cento per cento"}, title = {LT-VLM: Lithuanian Church Captioning Model (Gemma-3-4b-it LoRA)}, year = {2026}, howpublished = {\url{https://huggingface.co/AugusteBu/Cento_per_cento}}, }

APA:

Vilnius University Student Project Team "Cento per cento". (2026). LT-VLM: Lithuanian Church Captioning Model (Gemma-3-4b-it LoRA) [Model]. Hugging Face. https://huggingface.co/AugusteBu/Cento_per_cento

Glossary

LoRA: Low-Rank Adaptation – a parameter-efficient fine-tuning method.
VLM: Vision-Language Model – a model that processes both images and text.
Diacritics: Accent marks like ė, š, ž that are essential in Lithuanian orthography.

Model Card Authors

Vilnius University Student Project Team "Cento per cento"

Augustė Buklerytė
Aidas Jurelevičius
Karolina Mačiulevičiūtė

Model Card Contact

For questions or feedback, please open an issue in this repository or contact the team via databl.ai.

Framework versions

PEFT 0.19.1