Instructions to use dtiapkin/gemma3-4b-sft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dtiapkin/gemma3-4b-sft with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="dtiapkin/gemma3-4b-sft")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("dtiapkin/gemma3-4b-sft")
model = AutoModelForImageTextToText.from_pretrained("dtiapkin/gemma3-4b-sft")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use dtiapkin/gemma3-4b-sft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dtiapkin/gemma3-4b-sft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dtiapkin/gemma3-4b-sft",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/dtiapkin/gemma3-4b-sft

SGLang

How to use dtiapkin/gemma3-4b-sft with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "dtiapkin/gemma3-4b-sft" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dtiapkin/gemma3-4b-sft",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "dtiapkin/gemma3-4b-sft" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dtiapkin/gemma3-4b-sft",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use dtiapkin/gemma3-4b-sft with Docker Model Runner:
```
docker model run hf.co/dtiapkin/gemma3-4b-sft
```

See axolotl config

axolotl version: 0.13.0.dev0

# =========================
# Axolotl SFT config (Gemma3 4B PT, full finetune, bf16, grad ckpt)
# =========================

# ---- Model ----
base_model: google/gemma-3-4b-pt
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# Train only text generation (language model) layers
unfrozen_parameters:
  - "model.language_model.*"

trust_remote_code: true
strict: false

# Quantization OFF (full finetune)
load_in_8bit: false
load_in_4bit: false

# ---- Chat formatting ----
chat_template: gemma3

# ---- Dataset ----
# Option A: your dataset is already ShareGPT-style (list of turns with roles)
datasets:
  - path: RLHFlow/RLHFlow-SFT-Dataset-ver2
    type: chat_template
    field_messages: conversations
    roles_to_train: ["assistant"]
    split: train
    train_on_split: train


val_set_size: 0.01
train_on_inputs: false   # only learn on assistant tokens

# ---- Tokenization / packing ----
sequence_len: 8192        # <- start with 4096; later you can go 8192 once stable
sample_packing: true
pad_to_sequence_len: true

# Cache the prepared dataset (put on $WORK on Jean Zay)
dataset_prepared_path: ./prepared/gemma3-4b-8192
dataset_processes: 32 
dataloader_pin_memory: true
dataloader_num_workers: 8
dataloader_prefetch_factor: 2

# ---- Output / logging ----
output_dir:  ./outputs/gemma3-4b-sft
save_safetensors: true

logging_steps: 10
save_strategy: "epoch"
saves_per_epoch: 2
save_total_limit: 10

# Optional W&B
wandb_project: gemma3-sft
wandb_name: gemma3-4b-pt_seq8192_lr1.5e-5_bs128
wandb_watch:
wandb_log_model:

# ---- Precision / speed ----
bf16: true
fp16: false
tf32: true

flash_attention: true
xformers_attention:

# ---- Training hyperparams ----
num_epochs: 3                      # start with 1 epoch to validate end-to-end; then increase
micro_batch_size: 1
gradient_accumulation_steps: 16
auto_resume_from_checkpoints: True

optimizer: adamw_torch_fused       # simplest baseline (no bitsandbytes dependency surprises)
lr_scheduler: cosine
learning_rate: 1.5e-5              # good starting point for 12B full finetune
warmup_ratio: 0.05

weight_decay: 0.0
max_grad_norm: 1.0

group_by_length: false

# ---- Memory knobs ----
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
overrides_of_model_config:
  use_cache: false


# ---- Distributed ----
# When you launch with torchrun, Axolotl will use DDP.
# Keep these empty so you do NOT enable ZeRO/FSDP.
ddp:
deepspeed:
fsdp:
fsdp_config:


# ---- Debug ----
debug:

lustre/fsn1/projects/rech/jth/uay69xj/sft_outputs/gemma3-4b-sft

This model is a fine-tuned version of google/gemma-3-4b-pt on the RLHFlow/RLHFlow-SFT-Dataset-ver2 dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1.5e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 16
total_train_batch_size: 128
total_eval_batch_size: 8
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 137
training_steps: 2742

Training results

Framework versions

Transformers 4.57.1
Pytorch 2.8.0
Datasets 4.4.1
Tokenizers 0.22.2

Downloads last month: 3

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for dtiapkin/gemma3-4b-sft

Base model

google/gemma-3-4b-pt

Finetuned

(288)

this model

dtiapkin
/

gemma3-4b-sft

lustre/fsn1/projects/rech/jth/uay69xj/sft_outputs/gemma3-4b-sft

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

Model tree for dtiapkin/gemma3-4b-sft

Dataset used to train dtiapkin/gemma3-4b-sft