---
license: other
base_model: luca0621/appgen-qwen3-thinktag-lr5e-7-ep1
tags:
  - qwen3-vl
  - ms-swift
  - appgen
  - mobile-agent
  - continue-sft
---

# AppGen Qwen3 ThinkTag Continue SFT

Continued SFT checkpoint trained from `luca0621/appgen-qwen3-thinktag-lr5e-7-ep1`.

Training summary:

- Framework: MS-SWIFT 4.1.3
- Dataset: AMEX-derived continue-SFT set, 38,709 examples
- Epochs: 1
- Final step: 2420
- GPUs: 2 x NVIDIA B200
- Learning rate: 5e-7
- Precision: bfloat16
- DeepSpeed: ZeRO-2
- W&B project: `continue_train_app_gen_b200`
- SWIFT model type: `qwen3_vl`

Final training loss reported by SWIFT: `0.2788`.