Quantization 4-bit (bitsandbytes) LoRA rank (r) 12 LoRA α 24 LoRA dropout 0 Training epochs 5 Learning rate 8 × 10−5 LR scheduler Cosine Warmup steps 10 Per-device batch size 1 Gradient accumulation steps 8 Effective batch size 8 Optimizer AdamW (8-bit) Weight decay 0.02 Max sequence length 2048 tokens Eval/save strategy Per epoch (best checkpoint)

Evaluation

Testing Data, Factors & Metrics

Testing Data

Test data was taken from the test split of [https://huggingface.co/datasets/kniskiukas/gmm/]. Evaluation was done after each epoch.

Results

The fine-tuned model demonstrates improvements in: • Domain knowledge (accurate recognition of Lithuanian public transport vehicles, route numbers, and common street signs) • Naturalness and grammatical correctness of Lithuanian text • Focus on task-relevant elements instead of generic image descriptions

Model Card Authors

Mykolas Marius Budrys, Augustas Kniška, Emil Duko

Model Card Contact

mykolas.budrys@mif.stud.vu.lt, emil.duko@mif.stud.vu.lt, augustas.kniska@mif.stud.vu.lt

PEFT 0.19.1

mykolas-alt/LTVLM

Model Card for Model ID

Model Details

Model Description

Uses

Direct Use

Training Details

Training Data

Training Procedure

Preprocessing

Training Hyperparameters