Model Card: a2ui_qwen3_30b_sft_0410_v1

1. Overview

This model version is a 30B SFT-trained A2UI model based on the Qwen3 30B A3B instruct checkpoint. It is trained on the 0410 v11_2 image-url-fixed SFT dataset and serves as the warm-start parent for the 0410 DPO branch.

2. Model Metadata

Field	Value
model_version_id	a2ui_qwen3_30b_sft_0410_v1
algorithm	sft
base_model	Qwen/Qwen3-30B-A3B-Instruct-2507
precision	bf16

3. Lineage

Field	Value
parent_model_version_id	none

4. Training Data

Field	Value
dataset_version_id	a2ui_data_sft_0410_v1

5. Training Configuration

Field	Value
epochs_or_steps	epoch=1
batch_size	32
group_size	1
lr	1e-4
max_ctx	8192
max_resp	n/a
lora_rank	16

6. Change Summary

Train on the 0410 v11_2 image-url-fixed SFT dataset for 30B; this becomes the warm-start parent for the 0410 DPO branch.