hell0ks/gemma-4-26B-A4B-it-multitool

Gemma 4 26B A4B IT Multitool

An attempt to tool-friendly Gemma 4.

Problem was simple: The model knows how to use the functions and it does use it, but in lazy way.

1. It prioritize its knowledge than search tool.
2. It does not like to call multiple tools for one task.

So, I thought it might good chance to do some fine-tune, and I did.

The result was somewhat mixed bag:

• Good thing: It gets attracted to use search functions now, even with multiple calls!
• Bad thing: It thinks longer than before when using tools.

I think it's acceptable. That's why I upload this finetune.

SFT Data Source

Item	Details
Base model	trohrbaugh/gemma-4-26B-A4B-it-heretic-ara
Training method	LoRA SFT
Training library	Unsloth Studio
Source dataset	stepfun-ai/Step-3.5-Flash-SFT
Training samples	1,500

Dataset preparation

1. Normalize messages, tool definitions, arguments, and call identifiers.
2. Remove duplicate samples.
3. Check message structure and argument schemas.
4. Apply model-assisted quality review. (GPT-6 Luna, Low)

SFT Data Mix

Target action	Share	Training samples
First tool call	20%	300
Further call after a tool result	50%	750
Final answer or task completion	20%	300
No tool call or clarification	10%	150
Total	100%	1,500

Benchmarks

Tool-Eval-Bench

All models used FP8 inference and identical sampling parameters. Percentages specify the merge strength of the same trained LoRA adapter.

Model	Score /100	Raw points /136	Reasoning limit failures
Original	83	113	0
LoRA 100%	72	98	10
LoRA 50%	84	114	2
LoRA 75%	83	113	3

This is LoRA 75% checkpoint. Reason: 100% one was too aggressive and 50% one was still lazy.

License

It's Apache 2.0 and CC-BY-NC 2.0, followed by used dataseet.