Task

Given an image, the model generates a short Lithuanian caption focused on:

• cloudiness
• sunlight
• precipitation
• visibility
• time of day
• general weather conditions

Base model

unsloth/gemma-3-4b-it

Dataset

Dataset: Matas5/GMM_team_task

Training examples used: 103

The dataset contains images with Lithuanian weather captions. Captions were standardized to focus mainly on weather conditions rather than unrelated objects.

Training setup

Training method: LoRA fine-tuning with Unsloth

Model loading: 4-bit quantized base model

Fine-tuning type: PEFT / LoRA adapter

Run version: C

Purpose of this run: slower learning rate for more stable outputs

Number of epochs: 3

Total training steps: approximately 78

Per-device batch size: 1

Gradient accumulation steps: 2

Number of GPUs: 2 Tesla T4 GPUs

Effective total batch size: 4

Learning rate: 1e-4

Optimizer: adamw_8bit

Gradient checkpointing: enabled

Save strategy: save every epoch

LoRA rank: 4

LoRA alpha: 8

Target modules: q_proj, v_proj

Vision layers fine-tuned: yes

Prompt used during training

text
Copy code
Trumpai apibūdink orą šioje nuotraukoje lietuviškai. Atsakyk vienu paprastu sakiniu.

Example expected output style

text
Copy code
Dangus giedras ir ryškiai mėlynas, debesų beveik nėra. Oras saulėtas, sausas, matomumas labai geras.

Training result summary

Run C was trained for 3 epochs, like Run A, but with a smaller learning rate.

Run A used learning rate 2e-4.

Run B used 2 epochs with learning rate 2e-4.

Run C uses 3 epochs with learning rate 1e-4.

The purpose of this run is to test whether slower parameter updates produce more natural Lithuanian text and reduce strange or unstable phrases.

Because the dataset contains only 103 training examples, the result still needs to be evaluated on unseen test images.

Intended use

This adapter is intended for a university project demonstrating fine-tuning of a vision-language model for Lithuanian weather captioning.

It is not intended for professional meteorological forecasting.