@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.011_osd.yaml

@@ -0,0 +1,206 @@
+schema_version: 1
+id: aisb.t3.011_osd
+name: Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection
+version: 0.1.0
+one_line: Train and evaluate a CLIP-ViT-L/14 detector that uses SVD-based orthogonal
+  subspace decomposition (Effort) to generalize across deepfake and synthetic-image
+  forgery families on DeepfakeBench and GenImage benchmarks.
+task_description: 'This benchmark packages the "Effort" method for AI-generated image
+  detection. The core idea is to decompose CLIP-ViT-L/14 weight matrices via SVD into
+  frozen principal components (preserving pre-trained knowledge) and trainable residual
+  components (learning forgery patterns), enforcing orthogonality to prevent overfitting
+  to seen forgery types. The execution route involves: (1) downloading the CLIP-ViT-L/14
+  backbone and pre-trained Effort checkpoints from Google Drive, (2) obtaining deepfake
+  datasets externally from DeepfakeBench (FaceForensics++, Celeb-DF-v2, DF40) or GenImage/Chameleon
+  for synthetic-image evaluation, (3) training the Effort detector head on one source
+  dataset while keeping SVD-principal components frozen, and (4) evaluating cross-dataset
+  generalization accuracy across multiple unseen forgery methods. Two codebases are
+  bundled: DeepfakeBench (face deepfake detection, Table 1 results) and UniversalFakeDetect_Benchmark
+  (general synthetic image detection, Table 2 results). The primary metric is detection
+  accuracy on unseen forgery families (e.g., BigGAN, CycleGAN, SimSwap). Datasets
+  must be downloaded separately from external sources and are large (tens to hundreds
+  of GB). Checkpoints are hosted on Google Drive.
+
+  '
+capability_tags:
+- research_code_optimization
+- computer_vision
+- deepfake_detection
+- image_forensics
+- benchmark_evaluation
+- clip_finetuning
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 1d+
+difficulty: hard
+data_access: public
+primary_outputs:
+- biggan_accuracy
+- detection_scores
+- benchmark_report
+- cross_dataset_auc
+launch_profiles:
+- id: quick_check
+  label: Quick Check (Inference Only)
+  description: 'Download a pre-trained Effort checkpoint from Google Drive and run
+    DeepfakeBench/training/demo.py on a single image or folder to verify the inference
+    pipeline produces fake-probability scores.
+
+    '
+- id: single_dataset_train
+  label: Single-Dataset Train + Test
+  description: 'Train the Effort detector on one source dataset (e.g., FaceForensics++)
+    using a single GPU, then test on one held-out dataset (e.g., Celeb-DF-v2). Verifies
+    the training loop, SVD decomposition, orthogonal/KSV losses, and checkpoint saving.
+
+    '
+- id: cross_benchmark_eval
+  label: Cross-Benchmark Evaluation
+  description: 'Evaluate the trained or pre-trained Effort detector across multiple
+    unseen forgery datasets (SimSwap, BlendFace, DeepFaceLab, etc. via DF40) to reproduce
+    the cross-dataset generalization results from the paper.
+
+    '
+- id: univfd_benchmark
+  label: UniversalFakeDetect Benchmark
+  description: 'Use the UniversalFakeDetect_Benchmark/ codebase to reproduce Table
+    2 results on general synthetic image detection (GenImage, Chameleon).
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: external_repo
+    url: https://github.com/SCLBD/DeepfakeBench
+    access: public
+    note: 'FaceForensics++, Celeb-DF-v2, and other face deepfake datasets. Pre-processed
+      frames and face crops available via DeepfakeBench.
+
+      '
+  - kind: external_repo
+    url: https://github.com/YZY-stack/DF40
+    access: public
+    note: 'DF40 dataset with 40 distinct forgery methods for broader evaluation.
+
+      '
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/19kQwGDjF18uk78EnnypxxOLaG4Aa4v1h?usp=sharing
+    access: public
+    note: Pre-trained Effort checkpoints (FF++, GenImage, Chameleon).
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/1fm3Jd8lFMiSP1qgdmsxfqlJZGpr_bXsx?usp=drive_link
+    access: public
+    note: CLIP-ViT-L/14 backbone weights required by effort_detector.py.
+  notes:
+  - Deepfake datasets are tens to hundreds of GB after extraction.
+  - GenImage and Chameleon datasets must be obtained separately for Table 2 reproduction.
+  - JSON index files for each dataset must be generated or obtained from DeepfakeBench/DF40.
+  - CLIP backbone path is hardcoded in effort_detector.py and must be adjusted to
+    local path.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All datasets and checkpoints are publicly available but hosted externally on Google
+    Drive and GitHub repos.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 300
+    gpu_count: 4
+    gpu_vram_gb: 24
+environment:
+  python: 3.7.2
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - transformers
+  - loralib
+  - dlib==19.24.0
+  - opencv-python==4.6.0.66
+  - scikit-learn==1.0.2
+  - albumentations==1.1.0
+  - timm==0.6.12
+  - efficientnet-pytorch==0.7.1
+  - openai-clip (from git)
+  - kornia
+  - einops
+  notes:
+  - README specifies Python 3.7.2 and PyTorch 1.11, but install.sh installs torch==1.12.0+cu113;
+    use install.sh as the authoritative route.
+  - install.sh pins numpy==1.21.5, scipy==1.7.3 and many other specific versions.
+  - CLIP backbone uses HuggingFace transformers CLIPModel, not the OpenAI CLIP package
+    directly for the detector, but OpenAI CLIP is also installed.
+  - dlib is required for face landmark extraction during face-deepfake preprocessing.
+risk_flags:
+- external_dataset_dependency
+- external_checkpoint_dependency
+- hardcoded_paths
+- large_dataset_route
+- google_drive_download
+risk_notes:
+- Datasets (FaceForensics++, DF40, GenImage, Chameleon) are not bundled and must be
+  downloaded from external sources totaling potentially hundreds of GB.
+- CLIP backbone path is hardcoded as "../models--openai--clip-vit-large-patch14" in
+  effort_detector.py and must be manually adjusted.
+- Pre-trained checkpoints are hosted on Google Drive, which may rate-limit or require
+  manual browser download.
+- The benchmark was statically audited but no runtime execution was performed during
+  packaging; metric values are unverified.
+- install.sh installs torch 1.12 with CUDA 11.3 which may conflict with newer driver
+  setups.
+- JSON dataset index files must be generated via preprocessing scripts or obtained
+  from DeepfakeBench/DF40 repos.
+recommended_when: 'Use this benchmark when you want a realistic image-forensics optimization
+  task involving CLIP-based fine-tuning with SVD orthogonal decomposition, cross-dataset
+  generalization evaluation across many forgery families, and large-scale vision datasets.
+  Suitable for research into parameter-efficient tuning, deepfake detection robustness,
+  and foundation-model adaptation.
+
+  '
+not_recommended_when: 'Do not use this if you cannot provision 150+ GB of disk for
+  datasets and checkpoints, lack a CUDA GPU with at least 16 GB VRAM, need a lightweight
+  CPU-only benchmark, or cannot download large files from Google Drive and external
+  GitHub repositories.
+
+  '
+paper:
+  title: 'Effort: Efficient Orthogonal Modeling for Generalizable AI-Generated Image
+    Detection'
+  venue: ICML 2025 Oral
+  year: 2025
+  url: https://arxiv.org/abs/2411.15633
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.011_osd.zip
+  archive_type: zip
+  local_dir_name: paper-11-OSD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.011_osd.zip
+  sha256: db6c786efc3503420d8269cff59b8abd9ac07786364bc130a811615db99364d8
+  size_bytes: 29719270
+commercial:
+  annual_fee: null
+display:
+  palette_seed: obsidian-coral-vision
+  art_style: forensic-clean
+  accent_priority: high
+image_path: ../image/011_aisb.t3.011_osd.jpg

package/AISB/catalog/aisb.t3.011_osd.zh.yaml

@@ -0,0 +1,163 @@
+schema_version: 1
+id: aisb.t3.011_osd
+name: 正交子空间分解用于可泛化的AI生成图像检测
+version: 0.1.0
+one_line: 训练并评估一个使用基于SVD的正交子空间分解（Effort）的CLIP-ViT-L/14检测器，以在DeepfakeBench和GenImage基准上跨深度伪造和合成图像伪造家族实现泛化。
+task_description: '本基准将"Effort"方法打包用于AI生成图像检测。核心思想是通过SVD将CLIP-ViT-L/14权重矩阵分解为冻结的主成分（保留预训练知识）和可训练的残差成分（学习伪造模式），并强制正交性以防止对已知伪造类型的过拟合。执行流程包括：（1）从Google Drive下载CLIP-ViT-L/14主干网络和预训练的Effort检查点，（2）从外部获取深度伪造数据集（DeepfakeBench的FaceForensics++、Celeb-DF-v2、DF40）或GenImage/Chameleon用于合成图像评估，（3）在保持SVD主成分冻结的同时，在一个源数据集上训练Effort检测器头部，（4）在多个未见过的伪造方法上评估跨数据集泛化准确率。打包了两个代码库：DeepfakeBench（面部深度伪造检测，表1结果）和UniversalFakeDetect_Benchmark（通用合成图像检测，表2结果）。主要指标是在未见伪造家族上的检测准确率（如BigGAN、CycleGAN、SimSwap）。数据集必须从外部来源单独下载，且体积较大（数十到数百GB）。检查点托管在Google Drive上。
+
+  '
+capability_tags:
+- research_code_optimization
+- computer_vision
+- deepfake_detection
+- image_forensics
+- benchmark_evaluation
+- clip_finetuning
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 1d+
+difficulty: hard
+data_access: public
+primary_outputs:
+- biggan_accuracy
+- detection_scores
+- benchmark_report
+- cross_dataset_auc
+launch_profiles:
+- id: quick_check
+  label: 快速检查（仅推理）
+  description: '从Google Drive下载预训练的Effort检查点，并在单个图像或文件夹上运行DeepfakeBench/training/demo.py，以验证推理流程是否产生伪造概率分数。
+
+    '
+- id: single_dataset_train
+  label: 单数据集训练+测试
+  description: '使用单个GPU在一个源数据集（如FaceForensics++）上训练Effort检测器，然后在一个保留数据集（如Celeb-DF-v2）上进行测试。验证训练循环、SVD分解、正交/KSV损失和检查点保存功能。
+
+    '
+- id: cross_benchmark_eval
+  label: 跨基准评估
+  description: '在多个未见过的伪造数据集（通过DF40的SimSwap、BlendFace、DeepFaceLab等）上评估训练或预训练的Effort检测器，以复现论文中的跨数据集泛化结果。
+
+    '
+- id: univfd_benchmark
+  label: UniversalFakeDetect基准
+  description: '使用UniversalFakeDetect_Benchmark/代码库复现表2中关于通用合成图像检测（GenImage、Chameleon）的结果。
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: external_repo
+    url: https://github.com/SCLBD/DeepfakeBench
+    access: public
+    note: 'FaceForensics++、Celeb-DF-v2及其他面部深度伪造数据集。通过DeepfakeBench可获取预处理后的帧和人脸crop。
+
+      '
+  - kind: external_repo
+    url: https://github.com/YZY-stack/DF40
+    access: public
+    note: DF40数据集，包含40种不同的伪造方法，用于更广泛的评估。
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/19kQwGDjF18uk78EnnypxxOLaG4Aa4v1h?usp=sharing
+    access: public
+    note: 预训练的Effort检查点（FF++、GenImage、Chameleon）。
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/1fm3Jd8lFMiSP1qgdmsxfqlJZGpr_bXsx?usp=drive_link
+    access: public
+    note: effort_detector.py所需的CLIP-ViT-L/14主干网络权重。
+  notes:
+  - 深度伪造数据集解压后为数十到数百GB。
+  - GenImage和Chameleon数据集必须单独获取以复现表2。
+  - 每个数据集的JSON索引文件必须通过预处理脚本生成或从DeepfakeBench/DF40获取。
+  - CLIP主干路径在effort_detector.py中硬编码，必须调整为本地路径。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 所有数据集和检查点均可公开获取，但托管在外部的Google Drive和GitHub仓库上。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 300
+    gpu_count: 4
+    gpu_vram_gb: 24
+environment:
+  python: 3.7.2
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - transformers
+  - loralib
+  - dlib==19.24.0
+  - opencv-python==4.6.0.66
+  - scikit-learn==1.0.2
+  - albumentations==1.1.0
+  - timm==0.6.12
+  - efficientnet-pytorch==0.7.1
+  - openai-clip (from git)
+  - kornia
+  - einops
+  notes:
+  - README指定Python 3.7.2和PyTorch 1.11，但install.sh安装torch==1.12.0+cu113；请以install.sh作为权威安装路径。
+  - install.sh固定numpy==1.21.5、scipy==1.7.3及其他众多特定版本。
+  - CLIP主干使用HuggingFace transformers的CLIPModel，而非直接使用OpenAI CLIP包用于检测器，但也安装了OpenAI CLIP。
+  - dlib用于人脸深度伪造预处理期间的人脸特征点提取。
+risk_flags:
+- external_dataset_dependency
+- external_checkpoint_dependency
+- hardcoded_paths
+- large_dataset_route
+- google_drive_download
+risk_notes:
+- 数据集（FaceForensics++、DF40、GenImage、Chameleon）未打包，必须从外部来源下载，总量可能达数百GB。
+- CLIP主干路径在effort_detector.py中硬编码为"../models--openai--clip-vit-large-patch14"，必须手动调整。
+- 预训练检查点托管在Google Drive上，可能会有速率限制或需要手动浏览器下载。
+- 该基准已进行静态审计，但在打包过程中未执行运行时验证；指标值未经验证。
+- install.sh安装torch 1.12配合CUDA 11.3，可能与更新版本的驱动设置冲突。
+- 必须通过预处理脚本生成JSON数据集索引文件，或从DeepfakeBench/DF40仓库获取。
+recommended_when: '当您需要一个涉及基于CLIP的微调与SVD正交分解的真实图像取证优化任务、跨多个伪造家族的数据集泛化评估，以及大规模视觉数据集时，请使用此基准。适用于参数高效微调、深度伪造检测鲁棒性和基础模型适应方面的研究。
+
+  '
+not_recommended_when: '如果您无法为数据集和检查点提供150GB以上的磁盘空间、缺乏至少16GB显存的CUDA GPU、需要轻量级纯CPU基准，或无法从Google Drive和外部GitHub仓库下载大文件，请勿使用此基准。
+
+  '
+paper:
+  title: 'Effort: Efficient Orthogonal Modeling for Generalizable AI-Generated Image
+    Detection'
+  venue: ICML 2025 Oral
+  year: 2025
+  url: https://arxiv.org/abs/2411.15633
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.011_osd.zip
+  archive_type: zip
+  local_dir_name: paper-11-OSD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.011_osd.zip
+  sha256: db6c786efc3503420d8269cff59b8abd9ac07786364bc130a811615db99364d8
+  size_bytes: 29719270
+commercial:
+  annual_fee: null
+display:
+  palette_seed: obsidian-coral-vision
+  art_style: forensic-clean
+  accent_priority: high
+image_path: ../image/011_aisb.t3.011_osd.jpg

package/AISB/catalog/aisb.t3.012_efficientqat.yaml

@@ -0,0 +1,206 @@
+schema_version: 1
+id: aisb.t3.012_efficientqat
+name: 'EfficientQAT: Efficient Quantization-Aware Training for Large Language Models'
+version: 0.1.0
+one_line: 'Two-phase low-bit quantization-aware training (Block-AP → E2E-QP) for Llama-2/3
+  and Mistral families (7B–70B), evaluated via WikiText-2 perplexity, C4 perplexity,
+  and zero-shot accuracy on five reasoning benchmarks.
+
+  '
+task_description: 'This benchmark implements EfficientQAT, a two-stage quantization-aware
+  training pipeline for large language models at 2/3/4-bit weight-only precision.
+  Stage 1 (Block-AP) trains all parameters—weights, step sizes, and zero points—block-by-block
+  using reconstruction loss on 4096 RedPajama samples (context length 2048). Stage
+  2 (E2E-QP) freezes the quantized integer weights and fine-tunes only the quantization
+  step sizes end-to-end on a target dataset (RedPajama or Alpaca, context length 4096).
+  The primary executable anchors are main_block_ap.py and main_e2e_qp.py, with supplementary
+  scale-calibration scripts (calibrate_scales.py, calibrate_scales_v2.py). Evaluation
+  measures WikiText-2 perplexity, C4 perplexity, and average zero-shot accuracy on
+  WinoGrande, PIQA, HellaSwag, ARC-Easy, and ARC-Challenge using lm-eval v0.4.2. Pre-quantized
+  checkpoints for many model/bit configurations are available on HuggingFace and can
+  be converted to GPTQ or BitBLAS formats via bundled transfer scripts. The paper
+  reports that 2-bit Llama-2-70B can be obtained on a single A100-80GB in 41 hours
+  with <3 points accuracy degradation. No external evaluation service is required;
+  all metrics are computed locally.
+
+  '
+capability_tags:
+- research_code_optimization
+- large_language_models
+- quantization
+- model_compression
+- efficient_inference
+- quantization_aware_training
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1d+
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- wikitext2_ppl
+- c4_ppl
+- quantized_checkpoint
+- avg_accuracy
+launch_profiles:
+- id: calibration_only
+  label: Calibration Only
+  description: 'Run scale calibration (calibrate_scales.py or calibrate_scales_v2.py)
+    on a pre-quantized checkpoint using WikiText-2 train data. Produces updated scale
+    parameters without full Block-AP or E2E-QP retraining. Fastest route; requires
+    a pre-quantized model on disk.
+
+    '
+- id: qat_eval
+  label: QAT + Eval
+  description: 'Run the full two-stage EfficientQAT pipeline: Block-AP block-wise
+    training via main_block_ap.py, followed by E2E-QP end-to-end step-size training
+    via main_e2e_qp.py, then evaluate WikiText-2/C4 perplexity and zero-shot accuracy
+    on five reasoning tasks. This is the paper-faithful route and requires significant
+    GPU time (hours to days depending on model size).
+
+    '
+- id: eval_pretrained
+  label: Evaluate Pre-Quantized
+  description: 'Download a pre-quantized EfficientQAT checkpoint from HuggingFace
+    and evaluate perplexity and zero-shot accuracy without any training. Uses main_block_ap.py
+    with --resume_quant.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T-Sample
+    access: public
+    note: 'RedPajama calibration data used for both Block-AP (4096 samples, seqlen
+      2048) and E2E-QP (4096 samples, seqlen 4096). Downloaded automatically by datautils_block.py.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/wikitext
+    access: public
+    note: 'WikiText-2 (wikitext-2-raw-v1) used for perplexity evaluation and scale
+      calibration. Downloaded automatically during eval.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/ChenMnZ
+    access: public
+    note: 'Pre-quantized model checkpoints for Llama-2 (7B/13B/70B), Llama-3 (8B/70B),
+      Llama-3-Instruct, and Mistral-Large in EQAT/GPTQ/BitBLAS formats. Individual
+      checkpoints range from ~2 GB to ~39 GB.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/tatsu-lab/alpaca
+    access: public
+    note: 'Alpaca instruction-tuning dataset, used as alternative E2E-QP training
+      data for instruction-tuning scenarios.
+
+      '
+  notes:
+  - Full-precision base models (e.g. Llama-2-70B at ~131 GB) must be downloaded separately
+    from their original HuggingFace repos if running Block-AP from scratch.
+  - Pre-quantized checkpoints are much smaller (e.g. w2g64 Llama-2-70B is ~20 GB).
+  - Total disk usage depends heavily on which model families and bit widths are targeted.
+credential_requirements:
+  mode: none
+  items:
+  - HuggingFace account may be needed for gated models (Llama-2, Llama-3) but not
+    for the EfficientQAT checkpoints themselves
+  notes:
+  - Meta Llama model weights require accepting a license on HuggingFace before download.
+  - Pre-quantized EfficientQAT checkpoints under ChenMnZ namespace are publicly accessible.
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 48
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 500
+    gpu_count: 2
+    gpu_vram_gb: 80
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: 2.2.2
+  flash_attn: null
+  key_packages:
+  - bitsandbytes==0.41.0
+  - transformers==4.40.1
+  - lm-eval==0.4.2
+  - accelerate
+  - datasets
+  notes:
+  - See the bundled requirements.txt for the full dependency set.
+  - CUDA toolkit version is not pinned in the repo; any version supporting PyTorch
+    2.2.2 should work.
+  - GPTQModel (v0.9.8 tested) is required only for model format transfer to GPTQ/BitBLAS;
+    not needed for core QAT or evaluation.
+  - bitsandbytes is used for the AdamW optimizer in E2E-QP training.
+risk_flags:
+- large_model_download
+- high_gpu_memory
+- long_training_time
+- gated_model_access
+risk_notes:
+- 70B model routes require a single A100-80GB (paper reports 41 hours for 2-bit Llama-2-70B
+  Block-AP + E2E-QP) or multi-GPU setups.
+- 7B model routes are feasible on a single 48GB GPU but still take multiple hours
+  for full Block-AP + E2E-QP.
+- Full-precision base model weights for Llama-2/3 are gated on HuggingFace and require
+  license acceptance.
+- The --off_load_to_disk flag in Block-AP can reduce CPU memory usage at the cost
+  of training speed.
+- No benchmark execution was performed during the packaging pass; runtime verification
+  is still needed.
+- E2E-QP memory for 2-bit Llama-2-70B is reported as 34.2 GB in the paper.
+recommended_when: 'Use this benchmark when you want an LLM-systems task centered on
+  quantization-aware training at aggressive bit widths (2–4 bit), with self-contained
+  training and evaluation pipelines covering multiple model scales from 7B to 70B.
+  Suitable for evaluating optimization strategies for low-bit QAT, comparing against
+  PTQ and Q-PEFT baselines, or producing deployable quantized checkpoints in GPTQ/BitBLAS
+  formats.
+
+  '
+not_recommended_when: 'Do not use this if you lack access to at least one GPU with
+  ≥48 GB VRAM, if you need sub-hour benchmark turnaround, or if your focus is on small
+  models (<7B parameters). Also not suitable if you need weight-activation quantization
+  benchmarks (this benchmark is weight-only quantization).
+
+  '
+paper:
+  title: 'EfficientQAT: Efficient Quantization-Aware Training for Large Language Models'
+  venue: ACL 2025 Main
+  year: 2025
+  url: https://arxiv.org/abs/2407.11062
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.012_efficientqat.zip
+  archive_type: zip
+  local_dir_name: paper-12-EfficientQAT
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.012_efficientqat.zip
+  sha256: 8f53850f12f1bdbc4e3212b21fb51a7479ebb8ab715f3138d96ef1da979b2977
+  size_bytes: 107235
+commercial:
+  annual_fee: null
+display:
+  palette_seed: amber-steel-llm
+  art_style: hardware-editorial
+  accent_priority: high
+image_path: ../image/012_aisb.t3.012_efficientqat.jpg