iromu/Qwen2.5-1.5B-tools-NVFP4

Qwen2.5-1.5B Tools NVFP4

NVFP4 quantization of the Qwen2.5-1.5B-Instruct model fine-tuned for tool calling and agent-style interactions.

NVFP4 (FP4 e2m1 weights + FP8 e4m3 block scales, group 16) is the native precision on NVIDIA Blackwell and is served directly by TensorRT-LLM.

Base model

This model was fine-tuned from:

unsloth/Qwen2.5-1.5B-Instruct

Quantization

• —Quantization method: NVFP4 (W4A4), via NVIDIA ModelOpt 0.46.0 (mtq.NVFP4_DEFAULT_CFG), calibrated on 32 samples of the sft_tools train split
• —Export: modelopt.torch.export.export_hf_checkpoint

Training

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

LoRA configuration

• —LoRA dimension: 32
• —LoRA alpha: 32
• —Dropout: 0.05
• —Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training configuration

• —Max sequence length: 4096
• —Learning rate: 1e-5
• —Weight decay: 0.01
• —Global batch size: 64 (micro batch 2 x 32 accumulation)
• —Training steps: 501
• —Mixed precision: bf16

Dataset

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

Intended use

• —Structured tool/function calling
• —Agent-style multi-step interactions
• —Low-precision (FP4) serving on Blackwell GPUs

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

Usage

Serve the model with TensorRT-LLM:

trtllm-serve serve iromu/Qwen2.5-1.5B-tools-NVFP4 --port 8000

<!-- VALIDATION:BEGIN (auto-generated, do not edit) -->

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 (unsloth/Qwen2.5-1.5B-Instruct): 8.8% exact-args match (24/274). Fine-tuned (BF16): 0.7% exact-args match (2/274) (-8pp vs base).

<!-- VALIDATION:END -->