Qwen3-0.6B-Calculator-SFT

Calculator tool-use model fine-tuned from Qwen3-0.6B base.

Training

• Base Model: Qwen/Qwen3-0.6B
• Dataset: minpeter/calculator-sft (1,495 conversations)
• Framework: prime-rl SFT trainer

Hyperparameters

Parameter	Value
Learning Rate	1e-5
Batch Size	32
Micro Batch Size	8
Sequence Length	2048
Steps	50
Optimizer	AdamW

Evaluation

Metric	Score
Accuracy	99.2%
Avg Turns	2.0

Evaluated on 158 examples with 3 rollouts each using verifiers.

Usage

Chat Template

Uses standard ChatML format without thinking tokens.