@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.042_xpatch.zh.yaml

@@ -0,0 +1,182 @@
+id: aisb.t3.042_xpatch
+name: 'xPatch：基于指数季节-趋势分解的双流时间序列预测'
+version: 0.1.0
+one_line: 采用基于EMA的季节-趋势分解和长上下文补丁的CNN-MLP双流长期时间序列预测。
+task_description: '本封装基准测试包含xPatch，这是一个双流长期时间序列预测（LTSF）架构，使用指数移动平均（EMA）而非简单移动平均将输入序列分解为季节分量和趋势分量。线性流采用MLP层进行趋势预测，非线性流使用深度可分离CNN捕捉季节模式。该模型在非Transformer骨干网络中研究补丁化和通道独立技术，对标Transformer基线（PatchTST、Autoformer、FEDformer、CARD）和线性基线（DLinear、RLinear）在标准LTSF基准数据集上的表现。
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- forecast_checkpoints
+- ablation_outputs
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: '使用统一实验设置在单个准备好的预测数据集上运行xPatch。执行xPatch_unified脚本，使用默认超参数。适用于快速验证模型在单个基准数据集上的行为。
+
+    '
+- id: hyper_search
+  label: 超参数搜索
+  description: '在补丁长度、回溯窗口、预测长度和EMA alpha参数上运行全面的超参数搜索。使用xPatch_search脚本。生成的结果可与报告中使用drop-last技巧的基线论文进行公平比较。
+
+    '
+- id: fair_benchmark
+  label: 公平基准测试（TFB/OpenTS）
+  description: '遵循TFB基准测试协议，在不使用drop-last技巧的情况下执行公平实验。使用xPatch_fair脚本。结果可与OpenTS排行榜条目进行比较。
+
+    '
+- id: ablation_suite
+  label: 消融实验套件
+  description: '运行论文附录中的消融研究，包括EMA分解影响、arctangent损失函数效应、sigmoid学习率调整方案，以及与./ablation目录中基线模型的推理时间比较。
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: Google Drive
+    url: https://drive.google.com/u/0/uc?id=1NF7VEefXCmXuWNbnNe858WvQAkJ_7wuP&export=download
+    format: csv
+  - name: Baidu Pan
+    url: https://pan.baidu.com/share/init?surl=r3KhGd0Q9PJIUZdfEYoymg&pwd=i9iy
+    format: csv
+  - name: Kaggle Datasets
+    url: https://www.kaggle.com/datasets/wentixxiaogege/time-series-dataset
+    format: csv
+  notes:
+  - 将所有预处理数据集下载到./dataset目录
+  - 基准数据集包括：ETTh1、ETTh2、ETTm1、ETTm2、Weather、Traffic、Electricity
+  - 数据集已预处理完毕，可直接使用，无需额外标准化步骤
+credential_requirements:
+  mode: optional
+  items:
+  - kaggle_api_token
+  notes:
+  - Kaggle下载需要kaggle凭据或API令牌
+  - Google Drive和百度链接无需认证即可访问
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 单GPU足够；建议使用RTX 6000级或同等配置
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: '原始实验使用单张Quadro RTX 6000（24GB）。超参数搜索和消融研究受益于更大的内存，以便处理更大的批次。
+
+      '
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - pytorch>=2.0.0
+  - numpy
+  - pandas
+  - scikit-learn
+  - matplotlib
+  notes:
+  - 环境配置文件（environment.yml）指定了conda CUDA 11.8工具包，并使用pip安装PyTorch 2.5.1
+  - 混合conda/pip环境栈需要仔细的环境管理
+  - 使用以下命令创建conda环境：conda env create -f environment.yml
+  - 参见附带的README了解完整的依赖项集合
+  - ./scripts目录中的实验脚本处理运行时配置
+risk_flags:
+- mixed_environment_stack
+- data_download_external
+- hyperparameter_sensitivity
+risk_notes:
+- 混合conda CUDA 11.8与pip PyTorch 2.5.1可能导致版本冲突；请验证CUDA运行时与PyTorch内置CUDA版本匹配
+- 从Google Drive/Baidu下载外部数据集可能需要VPN或区域访问
+- 性能对补丁长度、回溯窗口、预测长度和EMA alpha参数敏感
+- 公平基准测试结果按照TFB协议排除了drop-last技巧；请验证哪种实验设置符合您的评估目标
+recommended_when: '当评估双流非Transformer预测架构、研究基于EMA的季节-趋势分解效果、探索Transformer之外的补丁化和通道独立技术，或在长期预测任务中将CNN/MLP方法与已建立的Transformer和线性基线进行比较时，使用此基准测试。
+
+  '
+not_recommended_when: '对于非时间序列任务、实时/在线预测场景、无法访问配备24GB显存的 workstation级GPU的环境，或需要内置不确定性量化的模型时，不建议使用此基准测试。
+
+  '
+paper:
+  title: 'xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend
+    Decomposition'
+  authors:
+  - Artyom Stitsyuk
+  - Jaesik Choi
+  venue: AAAI 2025
+  year: 2025
+  volume: '39'
+  number: '19'
+  pages: 20601-20609
+  url: https://ojs.aaai.org/index.php/AAAI/article/view/34270
+  arxiv_url: https://arxiv.org/pdf/2412.17323
+  bibtex: "@inproceedings{stitsyuk2025xpatch,\n  title={xPatch: Dual-Stream Time Series\
+    \ Forecasting with Exponential Seasonal-Trend Decomposition},\n  author={Stitsyuk,\
+    \ Artyom and Choi, Jaesik},\n  booktitle={Proceedings of the AAAI Conference on\
+    \ Artificial Intelligence},\n  volume={39}, number={19}, pages={20601--20609},\
+    \ year={2025}\n}\n"
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.042_xpatch.zip
+  archive_type: zip
+  local_dir_name: paper-42-xPatch
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.042_xpatch.zip
+  sha256: 72c8e6fd7e1a8b2eb184077ce9a37fb36514ef7bb148d8d0ff4ed79d4aa9de25
+  size_bytes: 3621184
+display:
+  palette_seed: teal-sunrise-trend
+  art_style: signal-geometry
+  accent_priority: high
+image_path: ../image/042_aisb.t3.042_xpatch.jpg
+metric_contract:
+  mse:
+    origin_path: utils/metrics.py
+    source_ref: MSE function
+    evaluation_protocol:
+      code_paths:
+      - run.py
+      - utils/metrics.py
+    status: code_backed
+  mae:
+    origin_path: utils/metrics.py
+    source_ref: MAE function
+    evaluation_protocol:
+      code_paths:
+      - run.py
+      - utils/metrics.py
+    status: code_backed
+  forecast_checkpoints:
+    origin_path: exp/exp_main.py
+    source_ref: model state saving
+    status: provisional
+code_anchors:
+- environment.yml
+- run.py
+- utils/metrics.py
+- exp/exp_main.py
+- layers/decomp.py
+- layers/network.py
+- layers/network_cnn.py
+- layers/network_mlp.py
+- models/xPatch.py
+caveat_snapshot: '静态代码审计发现utils/metrics.py和run.py中存在MSE和MAE指标的可执行锚点。本封装过程中未执行基准测试。运行时执行是信任指标值之前的必要步骤。公平基准测试需要TFB/OpenTS协议合规；请确认drop-last技巧处理与评估目标匹配。
+
+  '

package/AISB/catalog/aisb.t3.043_vhm.yaml

@@ -0,0 +1,268 @@
+id: aisb.t3.043_vhm
+name: VHM
+version: 0.1.0
+one_line: Two-stage remote-sensing VLM training (VersaD pretrain + instruction SFT)
+  with external RSEvalKit evaluation of a 7B model built on Vicuna-7B and CLIP-ViT-L-336px.
+task_description: 'This benchmark covers reproducing the full VHM (Versatile and Honest
+  Vision Language Model) pipeline for remote sensing image analysis. The model architecture
+  combines Vicuna-7B, CLIP-ViT-L/14-336px vision encoder, and a 2-layer MLP projector.
+  Training proceeds in two stages: (1) Pretrain on the VersaD dataset (1.4 million
+  RS image-text pairs with rich captions generated by Gemini-Vision), where layers
+  8+ of the vision encoder, the projector, and the full LLM are fine-tuned for one
+  epoch using DeepSpeed ZeRO-3 across multiple nodes; (2) Supervised fine-tuning (SFT)
+  on the VHM_SFT instruction dataset (covering VersaD-Instruct, HnstD honesty data,
+  and VariousRS-Instruct tasks including VQA, visual grounding, scene classification,
+  object counting, building vectorizing, and multi-label classification), where only
+  the projector and LLM are updated. Downstream evaluation requires cloning and installing
+  the external RSEvalKit repository in a separate conda environment, along with downloading
+  the VHM_Eval dataset (~7.35 GB). The packaged snapshot contains training code, DeepSpeed
+  configs, and SLURM launch scripts. Evaluation code, datasets, and model weights
+  are NOT bundled. No benchmark execution was performed during packaging; the primary
+  metric aid_accuracy is blocked pending RSEvalKit setup.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: very_high
+difficulty: hard
+data_access: public
+primary_outputs:
+- aid_accuracy
+- vhm_sft_checkpoint
+- rseval_report
+launch_profiles:
+- id: train_only
+  label: Train Only
+  description: 'Execute pretrain (slurm_pretrain.sh) and SFT (slurm_finetune.sh) stages
+    sequentially. Pretrain requires VersaD (~140 GB extracted), LLaVA-1.5 MLP projector
+    weights, Vicuna-7B, and CLIP-ViT-L-336px. SFT requires VHM_SFT (~124 GB extracted).
+    Produces pretrained and SFT checkpoints. Skips evaluation. Minimum viable on 2×48GB
+    GPUs via gradient accumulation; paper-reported timing assumes 16× A100-80GB pretrain
+    (~10h) and 8× A100-80GB SFT (~4h).
+
+    '
+- id: paper_faithful
+  label: Paper-Faithful
+  description: 'Run full pipeline: pretrain, SFT, then RSEvalKit evaluation. Requires
+    cloning RSEvalKit into a separate conda environment (rseval) and downloading VHM_Eval
+    (~7.35 GB). Produces all downstream metrics from the paper: AID scene classification
+    accuracy, RSVQA, DIOR-RSVG visual grounding, honest QA, building vectorizing,
+    and multi-label classification. Minimum 2×48GB GPUs for training; paper-reported
+    timing assumes 16× A100-80GB pretrain and 8× A100-80GB SFT. Honest QA evaluation
+    calls ChatGPT-3.5 API.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/datasets/FitzPC/VHM_VersaD
+    access: public
+    note: 'VersaD pretraining corpus: 1.4M RS image-text pairs with Gemini-Vision-generated
+      rich captions covering image properties, object attributes, and scene context.
+      Approximately 140 GB extracted. Sources include Million-AID (920K), CrowdAI/SpaceNet
+      (276K), fMoW (81K), CVUSA (44K), CVACT (44K), and LoveDA (24K). Resolution range
+      0.08–153 m/pixel.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/FitzPC/VHM_dataset_sft
+    access: public
+    note: 'VHM_SFT instruction dataset for supervised fine-tuning. Covers VersaD-Instruct,
+      HnstD factual and deceptive question pairs, and VariousRS-Instruct tasks. Approximately
+      124 GB extracted.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7B
+    access: public
+    note: 'Released VHM-7B SFT model weights. Allows skipping both training stages
+      and proceeding directly to RSEvalKit evaluation.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7b_pretrain_mlp_llm/tree/main
+    access: public
+    note: 'VHM pretrain-stage LLM and MLP projector weights. For resuming or initializing
+      from the pretrain checkpoint rather than training from scratch.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7b_pretrain_vit
+    access: public
+    note: 'CLIP ViT vision encoder weights after pretraining stage. Fine-tuned from
+      layer 8 onward during pretrain.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/liuhaotian/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5/tree/main
+    access: public
+    note: 'LLaVA-1.5 pretrained MLP projector (mm_projector.bin). Required initialization
+      for pretrain stage.
+
+      '
+  - kind: github_repo
+    url: https://github.com/fitzpchao/RSEvalKit
+    access: public
+    note: 'External evaluation toolkit for remote-sensing VLMs. Must be cloned and
+      installed in a dedicated conda environment (conda create -n rseval). All downstream
+      evaluation metrics are implemented here. VHM_Eval dataset download instructions
+      are in RSEvalKit docs/Data.md.
+
+      '
+  notes:
+  - 'Paper-faithful reproduction requires VersaD (~140 GB), VHM_SFT (~124 GB), VHM_Eval
+    (~7.35 GB), base model weights (Vicuna-7B, CLIP-ViT-L-336px), and LLaVA MLP projector.
+    Combined download and extracted footprint exceeds 400 GB before training checkpoints
+    and caches.
+
+    '
+  - 'VHM_Eval data is hosted externally; follow RSEvalKit docs/Data.md instructions
+    for download.
+
+    '
+  - 'All data and model weights are licensed for research use only (CC BY-NC 4.0).
+    Commercial use is prohibited. Model weights are further subject to LLaMA, Vicuna,
+    and Gemini license terms.
+
+    '
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 'No API keys or gated-access credentials are required for downloading any bundled
+    datasets or model weights.
+
+    '
+  - 'Honest QA evaluation in the paper calls the ChatGPT-3.5 API; reproducing this
+    exactly requires an OpenAI API key, but this call occurs inside RSEvalKit, not
+    in the training snapshot.
+
+    '
+resources:
+  minimum:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 400
+    gpu_count: 2
+    gpu_vram_gb: 48
+    gpu_arch: Ampere or newer
+  recommended:
+    cpu_cores: 64
+    ram_gb: 256
+    disk_gb: 1200
+    gpu_count: 16
+    gpu_vram_gb: 80
+    gpu_arch: A100
+    multi_node: true
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.0.1
+  flash_attn: 2.3.1.post1
+  key_packages:
+  - flash-attn==2.3.1.post1
+  - deepspeed==0.9.5
+  - transformers==4.31.0
+  - xformers
+  notes:
+  - 'requirements.txt pins torch 2.0.1+cu118, flash-attn 2.3.1.post1, and deepspeed
+    0.9.5. Create environment with: conda create -n vhm && conda activate vhm && pip
+    install -r requirements.txt.
+
+    '
+  - 'Training uses DeepSpeed ZeRO-3 (scripts/zero3.json) with bf16 mixed precision
+    and gradient checkpointing. Multi-node launch via torchrun with SLURM (scripts/rs/slurm_pretrain.sh).
+
+    '
+  - 'Three training entry points: vhm/train/train.py (base), vhm/train/train_mem.py
+    (FlashAttention, default), vhm/train/train_xformers.py (xformers attention, alternative).
+
+    '
+  - 'Evaluation requires a separate conda environment (rseval) with RSEvalKit requirements.txt.
+    A single environment is NOT sufficient for the paper-faithful route.
+
+    '
+  - 'Pretrain fine-tunes vision encoder from layer 8 onward, the MLP projector, and
+    the full LLM for one epoch. SFT fine-tunes only the projector and LLM.
+
+    '
+risk_flags:
+- route_caveat
+- external_eval_dependency
+- large_dataset_route
+- noncommercial_license
+- api_dependency
+risk_notes:
+- 'The paper-faithful route spans three disconnected steps: pretrain (slurm_pretrain.sh),
+  SFT (slurm_finetune.sh), and separate RSEvalKit evaluation in a different conda
+  environment—not one self-contained local command.
+
+  '
+- 'Paper-reported training times are 16× A100-80GB for pretrain (~10h) and 8× A100-80GB
+  for SFT (~4h). Slower runs on fewer or smaller GPUs should be budgeted in days.
+  Gradient accumulation steps can be adjusted to trade wall-clock time for reduced
+  GPU count.
+
+  '
+- 'Dataset preparation is substantial: VersaD ~140 GB, VHM_SFT ~124 GB, VHM_Eval ~7.35
+  GB, plus base model weights and generated checkpoints. Total disk footprint can
+  exceed 1 TB.
+
+  '
+- 'The primary evaluation metric (AID scene classification accuracy) depends on external
+  RSEvalKit, not a bundled evaluator. No executable evaluation anchor exists in the
+  current snapshot. The metric contract marks aid_accuracy as blocked.
+
+  '
+- 'Honest QA evaluation in the paper calls ChatGPT-3.5 API, introducing an external
+  API dependency and potential cost if reproduced exactly.
+
+  '
+- 'All data is CC BY-NC 4.0 (non-commercial only). Model use is further restricted
+  by LLaMA, Vicuna, and Gemini license terms. Research use only.
+
+  '
+recommended_when: 'Use this benchmark when you want to study remote-sensing VLM pretraining
+  on large-scale rich-caption datasets, investigate honesty in VLMs using factual/deceptive
+  question pairs, or reproduce the full two-stage training and evaluation pipeline.
+  Suitable for teams with multi-node GPU infrastructure (≥2 GPUs with ≥48GB VRAM each;
+  8–16× A100-80GB recommended), >400 GB free disk space, and willingness to set up
+  the external RSEvalKit evaluation environment.
+
+  '
+not_recommended_when: 'Do not use this for single-GPU smoke tests, quick evaluation-only
+  runs without training, or when you need a fully self-contained evaluator bundled
+  in the snapshot. Not suitable if you cannot tolerate multi-day training times, require
+  commercial-use licensing, lack multi-node GPU infrastructure, or need an executable
+  metric anchor without external dependencies.
+
+  '
+paper:
+  title: 'VHM: Versatile and Honest Vision Language Model for Remote Sensing Image
+    Analysis'
+  venue: AAAI 2025
+  year: 2025
+  url: https://arxiv.org/abs/2403.20213
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.043_vhm.zip
+  archive_type: zip
+  local_dir_name: paper-43-VHM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.043_vhm.zip
+  sha256: 73136beddf7624545dad64ba3818d9c7f5b78915ee52dfa157396f960ddd630f
+  size_bytes: 1902143
+commercial:
+  annual_fee: null
+display:
+  palette_seed: cobalt-emerald-orbit
+  art_style: satellite-atlas
+  accent_priority: high
+image_path: ../../../AISB/image/043_aisb.t3.043_vhm.jpg

package/AISB/catalog/aisb.t3.043_vhm.zh.yaml

@@ -0,0 +1,193 @@
+id: aisb.t3.043_vhm
+name: VHM
+version: 0.1.0
+one_line: 两阶段遥感视觉语言模型训练（VersaD预训练 + 指令SFT），使用外部RSEvalKit对基于Vicuna-7B和CLIP-ViT-L-336px构建的7B模型进行评估。
+task_description: '该基准测试涵盖复现完整的VHM（通用且诚实的视觉语言模型）遥感图像分析流程。模型架构结合了Vicuna-7B、CLIP-ViT-L/14-336px视觉编码器和2层MLP投影器。训练分两个阶段进行：（1）在VersaD数据集上进行预训练（140万对遥感图像-文本对，由Gemini-Vision生成丰富caption），其中对视觉编码器第8层及之后层、投影器和完整LLM进行一个epoch的微调，使用DeepSpeed ZeRO-3在多个节点上运行；（2）在VHM_SFT指令数据集上进行监督微调（SFT）（涵盖VersaD-Instruct、HnstD诚实数据和VariousRS-Instruct任务，包括VQA、视觉定位、场景分类、目标计数、建筑矢量化及多标签分类），仅更新投影器和LLM。下游评估需要在独立的conda环境中克隆并安装外部RSEvalKit仓库，同时下载VHM_Eval数据集（约7.35 GB）。打包的快照包含训练代码、DeepSpeed配置和SLURM启动脚本。评估代码、数据集和模型权重未包含在打包内。打包过程中未执行基准测试；主要指标aid_accuracy因待RSEvalKit设置而处于阻塞状态。
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: very_high
+difficulty: hard
+data_access: public
+primary_outputs:
+- aid_accuracy
+- vhm_sft_checkpoint
+- rseval_report
+launch_profiles:
+- id: train_only
+  label: 仅训练
+  description: '顺序执行预训练（slurm_pretrain.sh）和SFT（slurm_finetune.sh）阶段。预训练需要VersaD（约140 GB解压后）、LLaVA-1.5 MLP投影器权重、Vicuna-7B和CLIP-ViT-L-336px。SFT需要VHM_SFT（约124 GB解压后）。生成预训练和SFT检查点。跳过评估。最低可在2×48GB GPU上运行（通过梯度累积）；论文报告的时间基于16× A100-80GB预训练（约10小时）和8× A100-80GB SFT（约4小时）。
+
+    '
+- id: paper_faithful
+  label: 论文忠实复现
+  description: '运行完整流程：预训练、SFT，然后进行RSEvalKit评估。需要将RSEvalKit克隆到独立的conda环境中（rseval）并下载VHM_Eval（约7.35 GB）。生成论文中的所有下游指标：AID场景分类准确率、RSVQA、DIOR-RSVG视觉定位、诚实QA、建筑矢量化及多标签分类。训练最低需要2×48GB GPU；论文报告的时间基于16× A100-80GB预训练和8× A100-80GB SFT。诚实QA评估调用ChatGPT-3.5 API。
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/datasets/FitzPC/VHM_VersaD
+    access: public
+    note: 'VersaD预训练语料库：140万对遥感图像-文本对，包含Gemini-Vision生成的丰富caption，涵盖图像属性、目标特征和场景上下文。解压后约140 GB。来源包括Million-AID（92万）、CrowdAI/SpaceNet（27.6万）、fMoW（8.1万）、CVUSA（4.4万）、CVACT（4.4万）和LoveDA（2.4万）。分辨率范围0.08–153米/像素。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/FitzPC/VHM_dataset_sft
+    access: public
+    note: 'VHM_SFT监督微调指令数据集。涵盖VersaD-Instruct、HnstD事实性和欺骗性问答对以及VariousRS-Instruct任务。解压后约124 GB。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7B
+    access: public
+    note: '发布的VHM-7B SFT模型权重。可跳过两个训练阶段，直接进行RSEvalKit评估。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7b_pretrain_mlp_llm/tree/main
+    access: public
+    note: 'VHM预训练阶段的LLM和MLP投影器权重。用于从预训练检查点恢复或初始化，而非从头训练。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/FitzPC/vhm_7b_pretrain_vit
+    access: public
+    note: '预训练阶段后的CLIP ViT视觉编码器权重。预训练期间从第8层开始微调。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/liuhaotian/llava-v1.5-mlp2x-336px-pretrain-vicuna-7b-v1.5/tree/main
+    access: public
+    note: 'LLaVA-1.5预训练MLP投影器（mm_projector.bin）。预训练阶段必需初始化。
+
+      '
+  - kind: github_repo
+    url: https://github.com/fitzpchao/RSEvalKit
+    access: public
+    note: '遥感视觉语言模型外部评估工具包。必须克隆并在专用conda环境中安装（conda create -n rseval）。所有下游评估指标均在此实现。VHM_Eval数据集下载说明在RSEvalKit docs/Data.md中。
+
+      '
+  notes:
+  - '论文忠实复现需要VersaD（约140 GB）、VHM_SFT（约124 GB）、VHM_Eval（约7.35 GB）、基础模型权重（Vicuna-7B、CLIP-ViT-L-336px）和LLaVA MLP投影器。训练检查点和缓存前的综合下载和解压占用空间超过400 GB。
+
+    '
+  - 'VHM_Eval数据托管在外部；请按照RSEvalKit docs/Data.md中的说明进行下载。
+
+    '
+  - '所有数据和模型权重仅限研究使用（CC BY-NC 4.0）。禁止商业使用。模型权重还须遵守LLaMA、Vicuna和Gemini许可条款。
+
+    '
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - '下载任何打包的数据集或模型权重无需API密钥或受限访问凭证。
+
+    '
+  - '论文中的诚实QA评估调用ChatGPT-3.5 API；完全复现需要OpenAI API密钥，但该调用发生在RSEvalKit中，而非训练快照中。
+
+    '
+resources:
+  minimum:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 400
+    gpu_count: 2
+    gpu_vram_gb: 48
+    gpu_arch: Ampere or newer
+  recommended:
+    cpu_cores: 64
+    ram_gb: 256
+    disk_gb: 1200
+    gpu_count: 16
+    gpu_vram_gb: 80
+    gpu_arch: A100
+    multi_node: true
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.0.1
+  flash_attn: 2.3.1.post1
+  key_packages:
+  - flash-attn==2.3.1.post1
+  - deepspeed==0.9.5
+  - transformers==4.31.0
+  - xformers
+  notes:
+  - 'requirements.txt固定torch 2.0.1+cu118、flash-attn 2.3.1.post1和deepspeed 0.9.5。使用以下命令创建环境：conda create -n vhm && conda activate vhm && pip install -r requirements.txt。
+
+    '
+  - '训练使用DeepSpeed ZeRO-3（scripts/zero3.json），配合bf16混合精度和梯度检查点。通过torchrun和SLURM进行多节点启动（scripts/rs/slurm_pretrain.sh）。
+
+    '
+  - '三个训练入口点：vhm/train/train.py（基础）、vhm/train/train_mem.py（FlashAttention，默认）、vhm/train/train_xformers.py（xformers注意力，可选）。
+
+    '
+  - '评估需要独立的conda环境（rseval）和RSEvalKit requirements.txt。一个环境不足以支持论文忠实复现路线。
+
+    '
+  - '预训练对视觉编码器第8层及之后层、MLP投影器和完整LLM进行一个epoch的微调。SFT仅微调投影器和LLM。
+
+    '
+risk_flags:
+- route_caveat
+- external_eval_dependency
+- large_dataset_route
+- noncommercial_license
+- api_dependency
+risk_notes:
+- '论文忠实复现路线跨越三个独立步骤：预训练（slurm_pretrain.sh）、SFT（slurm_finetune.sh）和独立conda环境中的RSEvalKit评估——不是一个自包含的本地命令。
+
+  '
+- '论文报告的训练时间为16× A100-80GB预训练（约10小时）和8× A100-80GB SFT（约4小时）。在更少或更小的GPU上运行较慢，需要预留数天时间。可通过调整梯度累积步数来权衡时钟时间和减少的GPU数量。
+
+  '
+- '数据集准备量大：VersaD约140 GB、VHM_SFT约124 GB、VHM_Eval约7.35 GB，加上基础模型权重和生成的检查点。总磁盘占用可能超过1 TB。
+
+  '
+- '主要评估指标（AID场景分类准确率）依赖于外部RSEvalKit，而非打包的评估器。当前快照中不存在可执行的评估锚点。指标契约将aid_accuracy标记为阻塞状态。
+
+  '
+- '论文中的诚实QA评估调用ChatGPT-3.5 API，引入外部API依赖性和潜在成本（如需完全复现）。
+
+  '
+- '所有数据均为CC BY-NC 4.0（仅限非商业使用）。模型使用还受LLaMA、Vicuna和Gemini许可条款限制。仅限研究使用。
+
+  '
+recommended_when: '当您希望研究基于大规模丰富caption数据集的遥感视觉语言模型预训练、使用事实/欺骗性问答对研究视觉语言模型中的诚实性、或复现完整的两阶段训练和评估流程时使用此基准。适用于拥有多节点GPU基础设施（≥2块GPU，每块≥48GB VRAM；推荐8–16× A100-80GB）、超过400 GB可用磁盘空间且愿意设置外部RSEvalKit评估环境的团队。
+
+  '
+not_recommended_when: '不要将其用于单GPU冒烟测试、无训练的快速仅评估运行，或需要快照中完全自包含评估器的场景。不适用于无法容忍多日训练时间、需要商业使用许可、缺乏多节点GPU基础设施、或需要无外部依赖的可执行指标锚点的场景。
+
+  '
+paper:
+  title: 'VHM: Versatile and Honest Vision Language Model for Remote Sensing Image
+    Analysis'
+  venue: AAAI 2025
+  year: 2025
+  url: https://arxiv.org/abs/2403.20213
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.043_vhm.zip
+  archive_type: zip
+  local_dir_name: paper-43-VHM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.043_vhm.zip
+  sha256: 73136beddf7624545dad64ba3818d9c7f5b78915ee52dfa157396f960ddd630f
+  size_bytes: 1902143
+commercial:
+  annual_fee: null
+display:
+  palette_seed: cobalt-emerald-orbit
+  art_style: satellite-atlas
+  accent_priority: high
+image_path: ../../../AISB/image/043_aisb.t3.043_vhm.jpg