Training Results

This document summarizes the training results of our models enhanced with Verl-Tool across various agentic RL tasks.

Model Checkpoints

All these models are available in our Huggingface Collection.

Model	Link	Wandb
Qwen-2.5-Math-1.5B-Verl-tool	🤗	📈
Qwen-2.5-Math-7B-Verl-tool	🤗	📈

Model	Link	Wandb
VT-SQL-7B	🤗	📈

Model	Link	Wandb
VT-SWE-8B	🤗	📈

Model	Link	Wandb
VT-VisualReasoner-7B	🤗	📈

Model	Link	Wandb
VT-DeepSearch-8B	🤗	📈

1.5B Model Performance across challenging mathematical benchmarks:

Model Name	Tool	GSM8K	MATH 500	Minerva Math	Olympiad Bench	AIME24	AMC23	Avg
Qwen2.5-Math-1.5B	❌	39.50	34.80	8.10	23.00	13.30	35.00	25.62
Qwen2.5-Math-1.5B-Instruct	❌	84.90	74.20	26.80	39.00	10.00	57.50	48.70
Qwen2.5-Math-1.5B-Instruct + SimpleRL-Zoo	❌	81.90	70.20	20.60	33.90	20.00	55.00	46.90
Qwen-2.5-Math-1.5B-Instruct-TIR	✅	83.70	76.20	24.30	41.30	26.70	55.00	51.20
ToRL-1.5B	✅	85.60	77.80	29.80	44.00	26.70	67.50	55.23
Qwen-2.5-Math-1.5B + Verl-Tool	✅	85.10	77.40	28.30	44.00	33.30	65.00	55.52

7B Model Performance across challenging mathematical benchmarks:

Model Name	Tool	GSM8K	MATH 500	Minerva Math	Olympiad Bench	AIME24	AMC23	Avg
Qwen-2.5-Math-7B	❌	65.50	63.60	12.50	25.80	13.30	42.50	37.20
Qwen2.5-Math-7B-Instruct	❌	95.20	83.00	37.10	41.60	16.70	70.00	57.27
Qwen-2.5-Math-7B + SimpleRL-Zoo	❌	88.80	80.20	26.80	41.60	30.00	52.50	53.30
Qwen-2.5-Math-7B-Instruct-TIR	✅	94.60	82.40	29.00	50.50	30.00	62.50	58.17
TORL-7B	✅	92.70	82.20	33.50	49.90	43.30	65.00	61.10
Qwen-2.5-Math-7B + Verl-Tool	✅	91.40	83.40	29.80	50.20	40.00	72.50	61.22