---
title: Qwen3.5-4B-game-sft-lora-r64-step2000
canonical_url: "https://www.modelscope.cn/models/AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000"
md_url: "https://www.modelscope.cn/models/AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000.md"
repository: AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000
last_updated: 2026-10-23
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
base_model:
  - Qwen/Qwen3.5-4B
base_model_relation: adapter
parameters: 129.9M
tensor_type:
  - BF16
library_name:
  - lora
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 7
stars: 0
tags:
  - "base_model:adapter:Qwen/Qwen3.5-4B"
  - lora
  - transformers
  - multimodal
---

# Qwen3.5-4B-game-sft-lora-r64-step2000

> Qwen3.5-4B-game-sft-lora-r64-step2000 - AI-ModelScope 在 ModelScope 开源的模型。Qwen3.5-4B game trajectories SFT — step 2000

AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000 是 ModelScope 魔搭社区上的 129.9M 参数image-text-to-text模型，基于 Qwen/Qwen3.5-4B 构建。

- **Repository**: AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000
- **Tasks**: image-text-to-text
- **Parameters**: 129.9M
- **Base model**: Qwen/Qwen3.5-4B
- **Tags**: base_model:adapter:Qwen/Qwen3.5-4B, lora, transformers, multimodal
- **Downloads**: 7
- **Stars**: 0
- **Last updated**: 2026-10-23

Source: https://www.modelscope.cn/models/AI-ModelScope/Qwen3.5-4B-game-sft-lora-r64-step2000

---

# Qwen3.5-4B game trajectories SFT — step 2000

Deployable LoRA adapter saved at optimizer step 2000 while training on the
compressed `game-plus/trajectories-sft` multimodal trajectory format.

## Training configuration

- Base model: `Qwen/Qwen3.5-4B`
- PEFT: all-linear LoRA, rank 64, alpha 128, dropout 0.05
- Effective global batch size: 128
- Per-device micro batch size: 1 on 8 GPUs
- Gradient accumulation: 16
- Maximum sequence length: 32,768; overlength strategy: delete
- Maximum screenshots per sample: 4 (three prior plus current)
- Scheduler: cosine with 5% warmup
- Peak learning rate: 1e-4
- Precision: bfloat16
- ZeRO stage 1 and gradient checkpointing during training

This repository intentionally contains only the inference adapter and its
configuration. Optimizer, RNG, scheduler, and trainer resume states are not
included.