---
title: UHR-BAT2
canonical_url: "https://www.modelscope.cn/models/KAIWANG/UHR-BAT2"
md_url: "https://www.modelscope.cn/models/KAIWANG/UHR-BAT2.md"
repository: KAIWANG/UHR-BAT2
chinese_name: UHR-BAT2
last_updated: 2026-10-13
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen3_vl
architectures:
  - Qwen3VLForConditionalGeneration
base_model:
  - Qwen/Qwen3-VL-8B-Instruct
base_model_relation: finetune
parameters: 8.8B
tensor_type:
  - BF16
library_name:
  - pytorch
  - transformer
  - safetensors
frameworks:
  - pytorch
language:
  - en
  - zh
downloads: 5
stars: 0
tags:
  - qwen3_vl
  - vision-language
  - remote-sensing
  - visual-question-answering
  - geospatial
  - uhr-bat2
---

# UHR-BAT2

> UHR-BAT2 - KAIWANG 在 ModelScope 开源的模型。Qwen3-VL-8B fine-tuned for remote-sensing visual question answering. 遥感视觉问答微调模型。

KAIWANG/UHR-BAT2 是 ModelScope 魔搭社区上的 8.8B 参数image-text-to-text模型，采用 apache-2.0 许可，基于 Qwen/Qwen3-VL-8B-Instruct 构建。

- **Repository**: KAIWANG/UHR-BAT2
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 8.8B
- **Base model**: Qwen/Qwen3-VL-8B-Instruct
- **Tags**: qwen3_vl, vision-language, remote-sensing, visual-question-answering, geospatial, uhr-bat2
- **Downloads**: 5
- **Stars**: 0
- **Last updated**: 2026-10-13

Source: https://www.modelscope.cn/models/KAIWANG/UHR-BAT2

---

# UHR-BAT2

**A Qwen3-VL-8B model fine-tuned for remote-sensing visual question answering.**

**基于 Qwen3-VL-8B 微调的遥感视觉问答模型。**

[Hugging Face](https://huggingface.co/RL-MIND/UHR-BAT2) · [ModelScope / 魔搭](https://www.modelscope.cn/models/KAIWANG/UHR-BAT2) · [Evaluation details / 评测详情](EVALUATION.md)

UHR-BAT2 is a full-weight fine-tune of [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct), trained on a selected 10,000-example GeoLLaVA manifest. It keeps the native Qwen3-VL architecture and can be loaded with Transformers. The released checkpoint is `qwen3vl-8b-geollava-ft3-selected-10k`.

UHR-BAT2 在 Qwen3-VL-8B-Instruct 基础上，使用筛选后的 10,000 条 GeoLLaVA 样本进行微调，保留原生 Qwen3-VL 架构，可通过 Transformers 加载。发布检查点为 `qwen3vl-8b-geollava-ft3-selected-10k`，包含完整权重。

## Main results / 主要结果

These are complete, unpruned model evaluations, audited against existing prediction records on **2026-10-13**. The first three rows use the corrected option-scoring protocol. LRS-VQA uses generated-answer matching and is reported separately by metric.

以下为完整、不剪枝模型的评测结果，已于 **2026-10-13** 核对已有逐题预测记录。前三项使用修正后的选项计分协议；LRS-VQA 使用生成答案匹配，指标口径单独列明。

| Dataset / 数据集 | Accuracy / 准确率 | Correct / Total / 正确数与总数 | Metric / 指标 |
|---|---:|---:|---|
| **RSHR-Bench** | **50.72%** | 980 / 1,932 | All-QA option accuracy / 全部 QA 选项准确率 |
| **XLRS-Bench-lite** | **52.27%** | 1,610 / 3,080 | Strict option-set accuracy / 严格选项集合准确率 |
| **MME-RealWorld Remote Sensing** | **56.80%** | 2,123 / 3,738 | Full remote-sensing subset accuracy / 遥感全集准确率 |
| **LRS-VQA** | **23.78%** | 1,744 / 7,333 | Normalized generated-answer matching / 生成答案归一化匹配 |

RSHR's single-turn subset is **45.66% (726/1,590)**. XLRS includes **52.45% (1,563/2,980)** on single-choice questions and **47.00% (47/100)** on multi-select tasks. MME uses the full 3,738-record remote-sensing subset, not its 150-record Lite subset.

RSHR 单轮子集为 **45.66%（726/1,590）**。XLRS 单选题为 **52.45%（1,563/2,980）**，多选任务为 **47.00%（47/100）**。MME 使用完整 3,738 条遥感题，而非 150 条 Lite 子集。

The evaluated image budget is `min_pixels=784`, `max_pixels=1,310,720` per image. Images are resized within this budget. Single-choice evaluation uses FP32 candidate-option logits; XLRS multi-select evaluation uses constrained decoding and exact option-set matching. Ordinary free generation is not the same scoring procedure. See [EVALUATION.md](EVALUATION.md) and [evaluation_results.json](evaluation_results.json) for categories, denominators and protocol details.

评测中每张图的像素预算为 `min_pixels=784`、`max_pixels=1,310,720`，输入图像会按预算缩放。单选题采用 FP32 候选选项 logits；XLRS 多选题采用约束解码与严格选项集合匹配。普通自由生成与该评分流程不同。分类结果、分母及协议详见评测文档和 JSON。

## Model and training / 模型与训练

| Item / 项目 | Value / 配置 |
|---|---|
| Base model / 基座模型 | Qwen/Qwen3-VL-8B-Instruct |
| Architecture / 架构 | Qwen3VLForConditionalGeneration |
| Training data / 训练数据 | GeoLLaVA, selected 10K examples / 筛选 10K 样本 |
| Fine-tuning / 微调方式 | Frozen vision encoder; train merger and LLM / 冻结视觉编码器，训练 merger 与语言模型 |
| Checkpoint format / 权重格式 | Full BF16 safetensors, 4 shards / 完整 BF16 权重，4 个分片 |
| Training completion / 训练完成状态 | 1 epoch, 625 optimizer steps / 1 轮，625 个优化步骤 |
| Launch configuration / 启动配置 | Adafactor; learning rate 1e-5; global batch 16; maximum sequence length 8,192 |
| Release scope / 发布范围 | Dense checkpoint and inference example / 不剪枝检查点及推理示例 |

The checkpoint's trainer state confirms one completed epoch and 625 steps. Other training settings above are recorded launcher defaults. This is the final training checkpoint; no validation-selected best checkpoint was recorded. The original [UHR-BAT](https://huggingface.co/RL-MIND/UHR-BAT) uses a different architecture; its custom compression modules and results are not part of this Qwen3-VL weight release.

检查点训练状态确认完成 1 轮、625 步；表中其他训练参数来自保存的启动脚本默认配置。本次发布训练结束检查点，未记录通过验证集选优的检查点。原版 UHR-BAT 采用不同架构，其自定义压缩模块及成绩不属于本次 Qwen3-VL 权重发布。

## Quick start / 快速开始

Use Python 3.10 or later and a CUDA-enabled PyTorch installation. A single image can be processed with the standard Qwen3-VL implementation; custom remote code is not required. GPU memory requirements depend on image size and output length.

使用 Python 3.10 或更新版本及支持 CUDA 的 PyTorch。单图推理使用标准 Qwen3-VL 实现，无需自定义 remote code；显存需求随图像大小和输出长度变化。

```bash
pip install -r requirements.txt
python inference.py --model RL-MIND/UHR-BAT2 \
  --image /path/to/remote_sensing_image.jpg \
  --question "Describe the land use and visible objects in this image."
```

```python
import torch
from PIL import Image
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "RL-MIND/UHR-BAT2"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id, dtype="auto", device_map="auto", attn_implementation="sdpa"
).eval()

image = Image.open("remote_sensing_image.jpg").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Describe this remote-sensing image."},
]}]
prompt = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = processor(text=[prompt], images=[image], return_tensors="pt")
inputs.pop("token_type_ids", None)
inputs = inputs.to(model.device)
with torch.inference_mode():
    output = model.generate(
        **inputs, max_new_tokens=128, do_sample=False,
        temperature=None, top_p=None, top_k=None, use_cache=True,
    )
answer = output[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(answer, skip_special_tokens=True)[0])
```

For ModelScope, download the same package and pass the local directory to the example above:

通过魔搭下载相同模型包后，将本地目录传入上述示例：

```bash
pip install modelscope
```

```python
from modelscope import snapshot_download
model_id = snapshot_download("KAIWANG/UHR-BAT2")
# Use model_id with AutoProcessor and Qwen3VLForConditionalGeneration above.
```

The example demonstrates image-conditioned generation. It does not reproduce the option-scoring evaluation table by itself.

示例展示图像条件生成，本身不复现上表的候选选项计分流程。

The release passed a native loading and short image-generation smoke test with Python 3.11.0, PyTorch 2.5.1+cu121, Transformers 4.57.0.dev0 and Accelerate 1.13.0. The test checks package loading and generation, not benchmark accuracy or inference speed. File hashes and tensor metadata are in [release_manifest.json](release_manifest.json).

本次发布已通过原生加载与短图像生成测试，运行环境为 Python 3.11.0、PyTorch 2.5.1+cu121、Transformers 4.57.0.dev0 和 Accelerate 1.13.0。该测试验证模型包加载与生成，不代表准确率复测或速度测评。文件哈希及张量元数据见发布清单。

## Limitations / 局限

Counting, small-object localization and some open-answer categories remain challenging. These resized-input evaluations do not establish native full-resolution ultra-high-resolution performance. A matched evaluation of the unfine-tuned base model is not included, so this release makes no controlled fine-tuning improvement claim. A training-data audit found candidate scene-name/tile matches with RSHR; pixel-level and question-level overlap have not been resolved, so geographic-disjoint generalization is not established.

计数、小目标定位及部分开放回答类别仍存在不足。缩放输入评测不能证明原生完整超高分辨率性能；本次未提供同协议未微调基座对照，因此不宣称已验证的微调增益。训练数据审计发现与 RSHR 的场景名、瓦片候选匹配，尚未完成像素与问题层面的重合确认，因此未证明地理区域完全隔离条件下的泛化能力。

## License and acknowledgments / 许可与致谢

Released under [Apache-2.0](LICENSE), following the base model's license. See [NOTICE](NOTICE) for the upstream attribution and modifications. Training and benchmark datasets are not redistributed; their own terms apply. We acknowledge the Qwen team, GeoLLaVA contributors and the benchmark authors.

遵循基座模型许可，以 Apache-2.0 发布；上游归属及修改说明见 NOTICE。本仓库不重新分发训练集或评测集，数据仍适用各自条款。感谢 Qwen 团队、GeoLLaVA 贡献者及各基准作者。