Qwen2.5-1.5B-tools

Merged BF16 weights of the LoRA tool-calling fine-tune of unsloth/Qwen2.5-1.5B-Instruct. Built by scripts/build_tools_gguf.sh Qwen2.5-1.5B-tools (merge_and_unload of the latest checkpoint in checkpoints/Qwen2.5-1.5B-tools/).

Training

Recipe: TrainFinetuneRecipeForNextTokenPrediction (NeMo AutoModel 0.5.0)
Config: configs/qwen25_1.5b_tools.yaml
LoRA: dim 32, alpha 32, dropout 0.05, target_modules *.proj
Data: sft_tools split of r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation
Learning rate: 1e-5 (cosine decay), 501 steps, global batch 64, seq length 4096

Sibling artifacts

GGUF quantizations: models/Qwen2.5-1.5B-tools-GGUF/
NVFP4 checkpoint: models/Qwen2.5-1.5B-tools-NVFP4/

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).

GGUF-BF16: 2/274 (0.7%) exact, 88.6 tok/s — 100% of BF16.
GGUF-Q4_K_M: 3/274 (1.1%) exact, 142.9 tok/s — 150% of BF16.
GGUF-Q5_K_M: 6/274 (2.2%) exact, 131.4 tok/s — 300% of BF16.
GGUF-Q8_0: 2/274 (0.7%) exact, 115.3 tok/s — 100% of BF16.

Model	Quant	n	Tool call emitted	Names match	Exact args match	Δ exact vs BASE	tok/s
Qwen2.5-1.5B-tools	BASE (unsloth/Qwen2.5-1.5B-Instruct)	274	272/274 (99.3%)	59/274 (21.5%)	24/274 (8.8%)	—	40.6
Qwen2.5-1.5B-tools	BF16	274	258/274 (94.2%)	216/274 (78.8%)	2/274 (0.7%)	-8pp	38.0
Qwen2.5-1.5B-tools	GGUF-BF16	274	272/274 (99.3%)	167/274 (60.9%)	2/274 (0.7%)	-8pp	88.6
Qwen2.5-1.5B-tools	GGUF-Q4_K_M	274	248/274 (90.5%)	49/274 (17.9%)	3/274 (1.1%)	-7.7pp	142.9
Qwen2.5-1.5B-tools	GGUF-Q5_K_M	274	260/274 (94.9%)	79/274 (28.8%)	6/274 (2.2%)	-6.6pp	131.4
Qwen2.5-1.5B-tools	GGUF-Q8_0	274	271/274 (98.9%)	187/274 (68.2%)	2/274 (0.7%)	-8pp	115.3

iromu/qwen25-1.5b-tools

Qwen2.5-1.5B-tools

Training

Sibling artifacts

Validation matrix

iromu/
qwen25-1.5b-tools