pretty_name: Vision Encoder Evaluation Reproduction Data
language:
- en
task_categories:
- image-classification
- other
license: other
license_name: upstream-data-and-model-terms
license_link: >-
https://huggingface.co/datasets/336labs/VisionEncoder-Eval-ReproData/blob/main/LICENSES.md
size_categories:
- 100K<n<1M
source_datasets:
- liuhaotian/LLaVA-Pretrain
- ILSVRC/imagenet-1k
tags:
- vision-encoder-evaluation
- multimodal
- feature-extraction
- ravel
- numpy
- arxiv:2610.05413
configs:
- config_name: lcs-1k
default: true
data_files:
- split: train
path: lcs-1k/samples/samples.jsonl
- config_name: lcs-10k
data_files:
- split: train
path: lcs-10k/samples/samples.jsonl
- config_name: imagenet-200k
data_files:
- split: train
path: imagenet-200k/samples/samples.jsonl
A Strong Baseline for Evaluating Vision Encoders
in Multimodal Large Language Models
Yilin Yang1,* · Jun-Tao Tang2,* · Kengyi Wang3 · Siyuan Su3 · Gaoyong Luo4 · Mingda Chen1,†
1School of Artificial Intelligence, Shanghai Jiao Tong University
2Nanjing University ·
3Fudan University ·
4Independent Researcher
*Equal contribution. †Corresponding author.
Dataset Summary
Reproduction data, including precomputed vision and text features, for the evaluations in A Strong Baseline for Evaluating Vision Encoders in Multimodal Large Language Models. The release covers 70 vision encoders (43 language-supervised, 22 self-supervised, and 5 discrete), text features from three language models, and shared sample metadata.
Subsets
| Subset | Samples | Contents | Size |
|---|---|---|---|
lcs-1k |
1,000 | LCS-558K image patch features, text features, and captions | 55.3 GiB |
lcs-10k |
10,000 | LCS-558K image patch features, text features, and captions | 553.0 GiB |
imagenet-200k |
200,000 | ImageNet-1K pooled features, class labels, and kNN split | 59.9 GiB |
Each subset includes features for all 70 vision encoders. The LCS subsets also include text features from Qwen2.5-1.5B-Instruct, Qwen3-1.7B, and SmolLM2-1.7B-Instruct. The 10k arrays are stored in two 5k blocks per encoder/model.
Keep the shared sample row order when loading features. Shapes, extraction settings, and sampling details are recorded in the per-file metadata and sample manifests.
Encoder Catalog
encoders.json lists feature paths, shapes, metadata, and upstream encoder sources for all 70 encoders.
Data Sources
This repository provides the three cached feature subsets listed below. Raw images, MLLM training data, evaluation benchmarks, and other baseline datasets are obtained separately.
Data used by each evaluation methods
| Method | Data needed | Available here / source |
|---|---|---|
| RAVEL | Paired LCS-558K image patches and text features | lcs-1k/ and lcs-10k/ |
| RSA, CCA, GW, MutualNN | Paired LCS-558K global image vectors and text features | Shared samples and text are included; prepare the required global image vectors separately |
| kNN | ImageNet training features, labels, and support/query split | imagenet-200k/ |
| Linear probing | ImageNet train and validation images | ImageNet; the released 200k cache contains training images only |
| Alignment probing | COCO Karpathy captions; DreamLIP CC3M for the cross-dataset variants | COCO / Karpathy splits and DreamLIP captions |
| AC Policy / Law of Vision Representation | LCS-558K, Stage-1 projectors, and SPair-71k | LLaVA-Pretrain, Model Zoo, and SPair-71k |
| Mid-training loss | LCS-558K images/captions and projector-training logs | LLaVA-Pretrain; training instructions in the code repository |
| TokBench | Original images/annotations and matching tokenizer reconstructions | TokBench |
| MLLM training / evaluation | Training: LCS-558K and filtered mix665k; evaluation: the 11 benchmarks below | Model Zoo and the linked dataset sources |
Downstream MLLM benchmarks used in this paper
| Benchmark / required TSV | Source |
|---|---|
MMMU_TEST.tsv |
MMMU, test split |
MMBench_TEST_EN_V11.tsv |
MMBench, English test v1.1 |
VQAv2_VAL.tsv |
VQA v2, validation split |
ScienceQA_VAL.tsv |
ScienceQA, validation split |
ChartQA_TEST.tsv |
ChartQA, test split |
DocVQA_VAL.tsv |
DocVQA, validation split |
TextVQA_VAL.tsv |
TextVQA, validation split |
POPE.tsv |
POPE |
GQA_TestDev_Balanced.tsv |
GQA, balanced test-dev split |
MSCOCO_KARPATHY_TEST.tsv |
Karpathy splits and COCO images |
FLICKR30K_KARPATHY_TEST.tsv |
Karpathy splits and Flickr30k images |
Use the VLMEvalKit dataset loaders for benchmark preparation. Configuration used in our paper are in the code repository's Setup and Running sections.
Metadata and Integrity
The released arrays retain their original storage precision. Per-file metadata and sample manifests record available extraction and sampling information. FILES.json and checksums.sha256 provide file sizes and SHA-256 checksums.
Licensing and Attribution
See LICENSES.md. Use of the artifacts is subject to the applicable source-dataset and encoder terms. Source license information should be checked at the upstream links before reuse or redistribution; this card does not grant new rights over upstream images, captions, or model-derived artifacts.
Citation
If you use these feature caches or RAVEL, please cite the paper and the relevant source datasets and encoders:
@misc{yang2026strong,
title={A Strong Baseline for Evaluating Vision Encoders in Multimodal Large Language Models},
author={Yang, Yilin and Tang, Jun-Tao and Wang, Kengyi and Su, Siyuan and Luo, Gaoyong and Chen, Mingda},
year={2026},
eprint={2610.05413},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.05413}
}