tykhomyrov/quantgemma-checkpoints

QuantGemma Checkpoints

Fine-tuned Gemma 3 (270M) checkpoints from the quantgemma project.

Each checkpoint is a fine-tuned model saved after a training run. Folder names encode the key config.

Naming convention

{dataset}_bs{batch_size}_lr{lr}_steps{steps}/

Checkpoints

Path	val_loss	Dataset	Steps	Notes
v1-5m-full_bs8_lr2e-4_steps8612/	1.6367	v1-5m-full	8612	Best so far

Usage

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
    "tykhomyrov/quantgemma-checkpoints",
    subfolder="v1-5m-full_bs8_lr2e-4_steps8612",
    local_files_only=False,
)