Qwen2.5-1.5B k3

Qwen2.5-1.5B-Instruct fine-tuned with LoRA on Kimi-K3 distillation data for tool calling and agent-style interactions.

Base model

This model was fine-tuned from:

unsloth/Qwen2.5-1.5B-Instruct

Training

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

LoRA configuration

• LoRA dimension: 16
• LoRA alpha: 16
• 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: 2e-5
• Weight decay: 0.01
• Global batch size: 4 (micro batch 1 x 4 accumulation)
• Epochs: 1
• Mixed precision: bf16

Dataset

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

Intended use

• Structured tool/function calling
• Agent-style multi-step interactions
• Distilled reasoning behavior from the Kimi-K3 data mix

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

Usage

Serve the model:

trtllm-serve serve iromu/Qwen2.5-1.5B-k3 --port 8000

Load it with llama.cpp:

llama-cli -hf iromu/Qwen2.5-1.5B-k3-GGUF:Q4_K_M

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): 8.0% exact-args match (4/50). Fine-tuned (BF16): 0.7% exact-args match (2/274) (-7.3pp vs base).

• GGUF-BF16: 5/274 (1.8%) exact, 89.5 tok/s — 250% of BF16.
• GGUF-Q4_K_M: 7/274 (2.6%) exact, 128.6 tok/s — 350% of BF16.