数据集与方法

[glorialucheng]

SQL生成模型和SQL选择模型的训练方法以及数据集能否开源一下呢

[Peng-fei-Wang]

Hi @glorialucheng ,
Thanks for asking!

1. Data: We used the BIRD training set for the Generator, and the Selector's training data consists of candidates generated from that same set.
2. Training: Both models use the GRPO algorithm.
   Detailed hyperparameters and model configurations are provided in the table below:

Category	Hyperparameter	Generator	Selector
Data & Model	Base Model	OmniSQL-32B	Qwen2.5-Coder-32B
	Max Prompt Length	9216	9216
	Max Response Length	2048	2048
GRPO Algorithm	Reward Function	syntax, execution	selection
	Mini-Batch Size	64	128
	Micro-Batch Size / GPU	1	2
	KL Divergence Coefficient	0.001	0.001
	Learning Rate	1e-6	1e-6
	LR Warmup Steps Ratio	0.1	0.1
Rollout	Rollout Engine	SGLang	SGLang
	GPU Memory Utilization	0.5	0.8
	N	16	16
	Log-Prob Micro-Batch Size / GPU	1	2
	Ref Model Log-Prob Micro-Batch / GPU	1	2
	Tensor Parallel Size (TP)	8	8
Training	Total Epochs	3	5
	Total GPUs	32	32
	GPUs per Node	8	8
	Total Nodes	4	4
	GPU Type	A100	A100