Multi-task SFT on SQL (7,560) + JSON (3,568) + WebCode (1,107) = 12,235 examples combined. GSM8K flexible +20.5% over base Qwen2.5-Coder-3B (0.582→0.701). ARC stable.

Eval vs Baseline

Metric	Baseline	Multitask SFT	Delta
GSM8K flexible	0.5823	0.7013	+20.5%
GSM8K strict	0.6937	0.6907	-0.4%
ARC-acc	0.4556	0.4522	-0.7%
ARC-norm	0.4898	0.4949	+1.0%

Available Quants

File	Size	Use case
`*-Q2_K.gguf`	~1.5 GB	Minimum size, CPU inference
`*-Q3_K_M.gguf`	~1.9 GB	Small with decent quality
`*-Q4_K_M.gguf`	~2.2 GB	Recommended — best size/quality
`*-Q5_K_M.gguf`	~2.5 GB	High quality
`*-Q6_K.gguf`	~2.9 GB	Very high quality
`*-Q8_0.gguf`	~3.7 GB	Near-lossless

Usage (llama.cpp)

bash

llama-cli -m Qwen2.5-Coder-3B-SFT-StructuredOutput-Q4_K_M.gguf \
  -p "Write a SQL query to find all users who signed up in the last 30 days" \
  -n 256

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

Platform	Link
HuggingFace	huggingface.co/DuoNeural
Website	duoneural.com
GitHub	github.com/DuoNeural
X / Twitter	@DuoNeural

Subscribe: duoneural.beehiiv.com

DuoNeural/Qwen2.5-Coder-3B-SFT-StructuredOutput-GGUF

Qwen2.5-Coder-3B-SFT-StructuredOutput — GGUF

Eval vs Baseline

Available Quants

Usage (llama.cpp)

DuoNeural

DuoNeural/
Qwen2.5-Coder-3B-SFT-StructuredOutput-GGUF