You can deploy this model and start earning money today!

dikshant0520/
gemma-3-4b-it-mix-baseline4-lora

Name: gemma-3-4b-it-mix-baseline4-lora
Availability: InStock

Views

No views yet

Gemma 3 4B IT - Mixed Baseline 4 LoRA

LoRA adapter for google/gemma-3-4b-it, fine-tuned on 125,819 pseudo-labelled examples plus 2,000 gold examples. The adapter uses the language-model projection layers while leaving the vision tower frozen.

Training

Best checkpoint: step 13,000
Epochs: 2
LoRA rank/alpha/dropout: 16 / 32 / 0.10
Learning rate: 1e-4
Effective batch size: 15 across three GPUs
Best validation loss: 0.08334

Evaluation

On the shared 5,000-example evaluation set, mean LingoScore was 0.46035, median was 0.33312, and 35.12% of predictions scored at least 0.8.

Inference

Accept the Gemma license before downloading the base model. See inference.py --help. The input CSV requires question_id and question; the image parquet must contain the selected join column and an images column.

python inference.py --adapter dikshant0520/gemma-3-4b-it-mix-baseline4-lora --input-csv evaluation.csv --image-parquet images.parquet --output-csv predictions.csv