iromu/Gemma3-1B-tools-NVFP4

Gemma3 1B Tools NVFP4

The Gemma 3 1B tool-calling model (LoRA fine-tuned for tool calling and agent-style interactions) quantized to NVFP4 — FP4 e2m1 weights with FP8 e4m3 block scales (group 16), the Blackwell-native format.

Base model

This model was fine-tuned from:

google/gemma-3-1b-it

Quantization

Quantized from the merged BF16 checkpoint (iromu/Gemma3-1B-tools) with NVIDIA ModelOpt:

• Format: NVFP4 (W4A4) — NVFP4_DEFAULT_CFG, max scaling
• Calibration: 32 samples (batch 4, max 2048 tokens) from the sft_tools train split of r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation
• Target hardware: NVIDIA Blackwell (sm_100+), e.g. GB10 / DGX Spark

Training

Training was performed using NVIDIA NeMo AutoModel with LoRA/PEFT.

LoRA configuration

• LoRA dimension: 32
• LoRA alpha: 32
• Dropout: 0.05
• Target modules: *.proj (all *_proj linear layers)

Training configuration

• Max sequence length: 4096
• Learning rate: 5e-5 (cosine decay, 15 warmup steps, min 1e-6)
• Weight decay: 0.01
• Global batch size: 64 (micro batch 2 x 32 accumulation)
• Training steps: 336 (4 epochs)
• Mixed precision: bf16
• Validation loss: 0.579 → 0.4715 (final epoch)

Dataset

Training used the sft_tools split of the r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation dataset.

Tool-calling format

This model was trained with a custom chat template (bundled as chat_template.jinja). It renders the tool schemas into a developer turn and emits tool calls as:

<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>

Serving stacks must render prompts with this template (e.g. via the tokenizer's apply_chat_template) for tool calling to work.

Intended use

• Structured tool/function calling
• Agent-style multi-step interactions
• Small-footprint on-device or edge deployment

It is not intended to be a general replacement for larger Gemma models.

Usage

Serve the model with TensorRT-LLM:

trtllm-serve serve iromu/Gemma3-1B-tools-NVFP4 --port 8000

Validation matrix

Tool-calling validation on the sft_tools validation split (greedy decoding, 384 max new tokens). Throughput is single-stream greedy decode, not serving throughput.

Pretrained base (google/gemma-3-1b-it): 2.0% exact-args match (1/50). Fine-tuned (BF16): 66.0% exact-args match (33/50) (+64pp vs base).

• NVFP4: 2/50 (4.0%) exact, 169.0 tok/s — 6% of BF16.

Model	Quant	n	Tool call emitted	Names match	Exact args match	Δ exact vs BASE	tok/s
Gemma3-1B-tools	BASE (google/gemma-3-1b-it)	50	6/50 (12.0%)	1/50 (2.0%)	1/50 (2.0%)	—	68.5
Gemma3-1B-tools	BF16	50	50/50 (100.0%)	41/50 (82.0%)	33/50 (66.0%)	+64pp	47.1
Gemma3-1B-tools	NVFP4	50	34/50 (68.0%)	2/50 (4.0%)	2/50 (4.0%)	+2pp	169.0