Simplify subset descriptions and keep one download example

Summarize the three subsets in one table, retain one CLIP/Qwen3 download example, and fold the full encoder catalog and dataset-source details. Use a vision-and-text page title; preserve all feature files and dataset configs.

FILES.json

@@ -38,8 +38,8 @@
     },
     {
       "path": "README.md",
-      "bytes": 50075,
-      "sha256": "f157bc76e1cb382bb005a48bc637e81c5bf2a2af0ea9f2f7d0b3a50d2092e9fa"
+      "bytes": 37503,
+      "sha256": "1c1b8143eb2af87e8989b952879eea7648125b85e59284793e5dcff79f2f072d"
     },
     {
       "path": "encoders.json",
@@ -5026,7 +5026,7 @@
     }
   ],
   "total_files": 606,
-  "total_bytes": 717452462298,
+  "total_bytes": 717452449726,
   "source_lcs_manifest_sha256": "bc439048824c510b471d4c8cf34b143471a49a50260906f210cb8a1723e971ed",
   "published_lcs_manifest_sha256": "f934b05aadf9b94039b207812033e48b1e2aefe43e0fd084b5508e6fd103dbed",
   "validation": {
@@ -5049,5 +5049,5 @@
     "text_file_hashes_match_original_audits": true
   },
   "lcs10000_prepared_at": "2026-10-06T13:21:01.351852+00:00",
-  "documentation_updated_at_utc": "2026-10-07T04:03:11.305951+00:00"
+  "documentation_updated_at_utc": "2026-10-07T04:16:56.275968+00:00"
 }

README.md

@@ -1,5 +1,5 @@
 ---
-pretty_name: RAVEL Vision Encoder Features
+pretty_name: Vision and Text Evaluation Features
 language:
   - en
 task_categories:
@@ -38,7 +38,7 @@ configs:
 
 <h1 align="center">A Strong Baseline for Evaluating Vision Encoders<br>in Multimodal Large Language Models</h1>
 
-<p align="center"><strong>RAVEL · Vision Encoder Feature Dataset</strong></p>
+<p align="center"><strong>Precomputed Vision and Text Features</strong></p>
 
 <p align="center">Yilin Yang · Jun-Tao Tang · Kengyi Wang<br>Siyuan Su · Gaoyong Luo · Mingda Chen</p>
 
@@ -52,273 +52,63 @@ configs:
 
 ## Dataset Summary
 
-This dataset provides cached representations for the **70 vision encoders / visual tokenizers** evaluated in [A Strong Baseline for Evaluating Vision Encoders in Multimodal Large Language Models](https://arxiv.org/abs/2610.05413). It includes **43 language-supervised, 22 self-supervised, and 5 discrete** encoders, with shared sample order across encoders within each subset.
+Precomputed **vision and text features** for the evaluations in [A Strong Baseline for Evaluating Vision Encoders in Multimodal Large Language Models](https://arxiv.org/abs/2610.05413). 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.
 
-The release contains LCS-558K image patch features for RAVEL and other paired image–text probes, three matching language-model text feature arrays, and ImageNet-1K pooled image features for kNN and linear probes. Each encoder has its own metadata. The original NumPy arrays and storage precision are retained.
+**Resources:** [Paper](https://arxiv.org/abs/2610.05413) · [Code](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval) · [MLLM checkpoints and encoder weights](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo) · [Encoder catalog](encoders.json)
 
-**Release status:** **Complete.** The LCS-1000, LCS-10000, and ImageNet200k release file paths, sizes, and SHA-256 values have been verified against the Hub.
+## Subsets
 
-**Resources:** [Paper](https://arxiv.org/abs/2610.05413) · [Evaluation code](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval) · [MLLM Model Zoo and encoder weights](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo) · [Encoder catalog](encoders.json) · [File inventory](FILES.json)
+| Subset | Samples | Contents | Size |
+|---|---:|---|---:|
+| [`lcs-1k`](https://huggingface.co/datasets/336labs/VisionEncoder-Features/tree/main/lcs-1k) | 1,000 | LCS-558K image patch features, text features, and captions | 55.3 GiB |
+| [`lcs-10k`](https://huggingface.co/datasets/336labs/VisionEncoder-Features/tree/main/lcs-10k) | 10,000 | LCS-558K image patch features, text features, and captions | 553.0 GiB |
+| [`imagenet-200k`](https://huggingface.co/datasets/336labs/VisionEncoder-Features/tree/main/imagenet-200k) | 200,000 | ImageNet-1K pooled features, class labels, and kNN split | 59.9 GiB |
 
-<a id="datasets-required-by-the-evaluation-pipelines"></a>
-
-## Data requirements
-
-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 method
-
-| 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](https://www.image-net.org/download.php); the released 200k cache contains training images only |
-| Alignment probing | COCO Karpathy captions; DreamLIP CC3M for the cross-dataset variants | [COCO / Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [DreamLIP captions](https://huggingface.co/datasets/qidouxiong619/dreamlip_long_captions) |
-| AC Policy / Law of Vision Representation | LCS-558K, Stage-1 projectors, and SPair-71k | [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain), [Model Zoo](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo), and [SPair-71k](https://cvlab.postech.ac.kr/research/SPair-71k/index.html) |
-| Mid-training loss | LCS-558K images/captions and projector-training logs | [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain); training instructions in the [code repository](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#running) |
-| TokBench | Original images/annotations and matching tokenizer reconstructions | [TokBench](https://huggingface.co/datasets/Junfeng5/TokBench) |
-| MLLM training / evaluation | Training: LCS-558K and filtered mix665k; evaluation: the 11 benchmarks below | [Model Zoo](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo) and the linked dataset sources |
-
-ImageNet kNN features have class labels but no paired captions, so they cannot supply an alignment probe's image–text inputs on their own. A full linear-probe benchmark also needs independent validation data.
-
-### Raw data folders
-
-The following are example local folders. Keep the image subdirectories recorded in each dataset's annotations.
-
-```text
-data/
-  instructions/pretrain/blip_laion_cc_sbu_558k.json
-  instructions/finetune/llava_v1_5_mix665k_drop_ge8kchars.json
-  instructions/test/<dataset>.tsv
-  images/pretrain/
-  images/finetune/
-  images/test/
-  imagenet/{train,val}/
-  coco/
-  cc3m/
-  SPair-71k/
-  tokbench/
-```
-
-**MLLM pretraining:** obtain the JSON and `images.zip` from [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain). Image names in the JSON are relative to `data/images/pretrain/`.
-
-**MLLM finetuning:** obtain [mix665k](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_v1_5_mix665k.json) and its COCO, GQA, OCR-VQA, TextVQA, and Visual Genome images using the [LLaVA data preparation instructions](https://github.com/haotian-liu/LLaVA#visual-instruction-tuning). The project uses `llava_v1_5_mix665k_drop_ge8kchars.json`, which excludes 395 examples with at least 8,000 conversation-text characters. Image names in this JSON are relative to `data/images/finetune/`.
-
-### Downstream MLLM benchmarks
-
-Place the prepared benchmark TSVs in `data/instructions/test/`. Obtain the matching data from these sources:
+Each subset includes features for all 70 vision encoders. The LCS subsets also include text features from [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct), [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B), and [SmolLM2-1.7B-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct). The 10k arrays are stored in two 5k blocks per encoder/model.
 
-| Benchmark / required TSV | Source |
-|---|---|
-| `MMMU_TEST.tsv` | [MMMU](https://huggingface.co/datasets/MMMU/MMMU), test split |
-| `MMBench_TEST_EN_V11.tsv` | [MMBench](https://github.com/open-compass/MMBench), English test v1.1 |
-| `VQAv2_VAL.tsv` | [VQA v2](https://visualqa.org/download.html), validation split |
-| `ScienceQA_VAL.tsv` | [ScienceQA](https://github.com/lupantech/ScienceQA), validation split |
-| `ChartQA_TEST.tsv` | [ChartQA](https://github.com/vis-nlp/ChartQA), test split |
-| `DocVQA_VAL.tsv` | [DocVQA](https://www.docvqa.org/datasets/docvqa), validation split |
-| `TextVQA_VAL.tsv` | [TextVQA](https://textvqa.org/), validation split |
-| `POPE.tsv` | [POPE](https://github.com/RUCAIBox/POPE) |
-| `GQA_TestDev_Balanced.tsv` | [GQA](https://cs.stanford.edu/people/dorarad/gqa/download.html), balanced test-dev split |
-| `MSCOCO_KARPATHY_TEST.tsv` | [Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [COCO images](https://cocodataset.org/#download) |
-| `FLICKR30K_KARPATHY_TEST.tsv` | [Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [Flickr30k images](https://shannon.cs.illinois.edu/DenotationGraph/) |
-
-Prepare images referenced by the TSVs under `data/images/test/`. Folder names follow the benchmark loader: for example, MMBench v1.1 uses `MMBench_V11/`. TSVs with embedded images can be decoded by the loader. GQA also accepts the filename `GQA_TESTDEV_BALANCED.tsv`.
-
-Use the [VLMEvalKit dataset loaders](https://github.com/open-compass/VLMEvalKit/tree/main/vlmeval/dataset) for benchmark preparation. Original dataset downloads must be converted to the required TSV format; changing a filename alone is insufficient. Configuration and experiment commands are in the code repository's [Setup](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#setup) and [Running](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#running) sections.
-
-## Dataset Structure
-
-| Component | Arrays | Shape | Storage dtype | Size (GiB) |
-| --- | ---: | --- | --- | ---: |
-| LCS-1000 visual patches | 70 | `[1000, T, D]` | float16 | 55.27 |
-| Aligned LLM text vectors | 3 | `[1000, D]` | float32 | 0.02 |
-| LCS-10000 visual patches (two blocks per encoder) | 140 | `[5000, T, D]` | float16 | 552.74 |
-| LCS-10000 historical last-token text (two blocks per model) | 6 | `[5000, D]` | float32 | 0.21 |
-| ImageNet200k pooled features | 70 | `[200000, D]` | float32 | 59.89 |
-
-Total prepared contents, including shared metadata and protocols: approximately **668.18 GiB**. NumPy sizes include their headers; no extra precision conversion was applied.
+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.
 
 ```text
 features/
   lcs-1k/{vision,text,samples}/
   lcs-10k/{vision,text,samples}/
-  imagenet-200k/
-    features/
-    metadata/
-    samples/
-    protocols/
-    labels.npy
-    source_indices.npy
+  imagenet-200k/{features,metadata,samples,protocols}/
   encoders.json
   FILES.json
   checksums.sha256
 ```
 
-### LCS-1k: 1,000 paired examples
-
-The source is the `blip_laion_cc_sbu_558k.json` alignment annotation set from [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain). The cached subset was sampled with `numpy.default_rng(42).choice`, without replacement, retaining the **unsorted sampled order**. The ordered manifest identifies the exact selected records; the original full-source annotation hash/revision is not recorded in every historical audit.
-
-Each visual array has shape **`[1000, T, D]`**, stored as **float16**; patch count `T` and dimension `D` depend on the encoder. The text arrays are float32, with one vector per matching caption:
+## Download
 
-| ID | Language model | Shape | Text extraction |
-| --- | --- | --- | --- |
-| `qwen25` | [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) | `[1000, 1536]` | Penultimate hidden state, mean over valid tokens |
-| `qwen3` | [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) | `[1000, 2048]` | Penultimate hidden state, mean over valid tokens |
-| `smollm2` | [SmolLM2-1.7B-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct) | `[1000, 2048]` | Penultimate hidden state, mean over valid tokens |
-
-Text extraction uses layer `-2`, right padding, no truncation, bfloat16 inference, and float32 storage. The audits record pinned model revisions, file hashes, extraction settings, and software versions. These are the **mean-token** text features used by the current paper reproduction pipeline.
-
-Row `i` in **every visual array and every text array** refers to row `i` of `samples/samples.jsonl`. Keep this order when comparing encoders. Fields are:
-
-| Field | Meaning |
-| --- | --- |
-| `row_index` | Zero-based row in the released arrays |
-| `source_index` | Record index in the original LCS alignment annotation list |
-| `image_id` | Image path relative to the upstream image archive |
-| `text_id` | Source text/example identifier |
-| `text` | Caption used for text feature extraction |
-
-The full manifest additionally retains the source conversation records and sampling information. Raw images are available from the upstream dataset.
-
-### LCS-10k: 10,000 paired examples
-
-This larger cache contains **70 encoders × 10,000 images** as raw float16 patch tokens, stored in **two ordered 5,000-row blocks per encoder**. The first block uses the fixed seed-42 sample; the second uses seed 43 after excluding the first block by source annotation index, image ID, text ID, and image-file content hash. The original first block contains 11 repeated image-content rows; the disjoint-content check applies between blocks and within the new block. The two blocks have matching patch shapes and storage dtypes for each encoder.
-
-| Block | Combined rows | Visual array shape | Sample order |
-| --- | --- | --- | --- |
-| `part0_seed42` | `0:5000` | `[5000, T, D]` | Fixed seed-42 sample, unsorted |
-| `part1_seed43_disjoint` | `5000:10000` | `[5000, T, D]` | Disjoint seed-43 complement, unsorted |
-
-The combined sample table has **10,000 records**. `row_index` addresses the combined sample table; `block_index` and `block_row_index` address the stored feature array. `source_index` identifies the original LCS annotation record. The released `source_indices.npy` files are reconstructed from the canonical manifests. Historical `indices_n5000` / `indices_new5000` files contained only `0..4999` block row positions; those positions are not original annotation indices.
-
-The paired text cache contains **Qwen2.5-1.5B-Instruct, Qwen3-1.7B, and SmolLM2-1.7B-Instruct**, each as two float32 arrays of `[5000, D]` (`D=1536, 2048, 2048`). These historical arrays use the **penultimate hidden state at the last non-padding token**. Their metadata retains the original extraction audit and manifest hashes. Exact historical model revisions are absent from these audits.
-
-Use [feature_manifest.json](lcs-10k/feature_manifest.json), the block sample manifests, and [index_semantics.json](lcs-10k/samples/index_semantics.json) when loading the arrays. No visual pooling, PCA, whitening, or precision conversion was applied to these blocks.
-
-### ImageNet-200k: 200,000 training examples
-
-The subset contains **200 examples per class**, across **1,000 classes**, from the official ImageNet-1K training split. The export records a **support-sampling seed of 0**; the exact selected examples are given by `source_indices.npy`. All 70 arrays have shape **`[200000, D]`** and are stored as **float32**. Shared `source_indices.npy` is strictly increasing in the original train-index order; `labels.npy` contains integer class IDs `0–999`.
-
-The sample table contains `row_index`, `source_index`, and `label`. Each array's readout is recorded in its metadata: for example, CLIP uses its final post-LayerNorm CLS before the visual projection, while the DINO readout concatenates normalized CLS and mean patch features. **Use the per-encoder representation description when interpreting dimensions.** These pooled/readout arrays have different semantics from the LCS patch-token arrays.
-
-`protocols/knn_seed42.json` provides the existing deterministic **195-example training pool / 5-example query split per class**, with nested training subsets of **5, 10, 20, 45, 95, and 195** examples per class. Protocol indices address the **released 200,000-row arrays**; source indices identify examples in the original training split. The query split consists of held-out examples from that training subset. Full ImageNet training and validation features are outside this release.
-
-## Download and Use
-
-Download into `features/` and keep the repository subfolders. File paths in `encoders.json` and `FILES.json` are relative to this folder. Install `huggingface_hub` and `numpy`; install `datasets` to load the sample tables.
-
-### Download one encoder with aligned text features
+Example: download the CLIP ViT-L/14 image features and matching Qwen3 text features from `lcs-1k` into `features/`.
 
 ```python
-import json
-from pathlib import Path
-import numpy as np
 from huggingface_hub import snapshot_download
-
-root = Path(snapshot_download(
-    repo_id="336labs/VisionEncoder-Features",
-    repo_type="dataset",
-    allow_patterns=[
-        "encoders.json", "README.md", "LICENSES.md",
-        "lcs-1k/samples/**",
-        "lcs-1k/vision/clip_openai__l14*",
-        "lcs-1k/text/qwen3*",
-    ],
-    local_dir="features",
-))
-base = root / "lcs-1k"
-vision = np.load(base / "vision/clip_openai__l14_patch_n1000_seed42.npy",
-                 mmap_mode="r", allow_pickle=False)
-text = np.load(base / "text/qwen3_penultimate_mean_n1000_seed42.npy",
-               mmap_mode="r", allow_pickle=False)
-samples = json.loads((base / "samples/manifest.json").read_text())["records"]
-assert vision.shape[0] == text.shape[0] == len(samples) == 1000
-```
-
-### Download one encoder from the 10,000-sample cache
-
-```python
-import json
-from pathlib import Path
 import numpy as np
-from huggingface_hub import snapshot_download
 
-root = Path(snapshot_download(
+root = snapshot_download(
     repo_id="336labs/VisionEncoder-Features",
     repo_type="dataset",
-    allow_patterns=[
-        "encoders.json", "README.md", "LICENSES.md",
-        "lcs-10k/feature_manifest.json",
-        "lcs-10k/samples/**",
-        "lcs-10k/vision/clip_openai__l14/**",
-        "lcs-10k/text/qwen3/**",
-    ],
     local_dir="features",
-))
-base = root / "lcs-10k"
-blocks = ["part0_seed42", "part1_seed43_disjoint"]
-vision_parts = [np.load(base / f"vision/clip_openai__l14/{part}.npy",
-                        mmap_mode="r", allow_pickle=False) for part in blocks]
-text_parts = [np.load(base / f"text/qwen3/{part}_penultimate_lasttok.npy",
-                      mmap_mode="r", allow_pickle=False) for part in blocks]
-samples = json.loads((base / "samples/manifest.json").read_text())["records"]
-assert len(samples) == sum(x.shape[0] for x in vision_parts) == 10000
-for block_index in range(2):
-    assert vision_parts[block_index].shape[0] == text_parts[block_index].shape[0] == 5000
-# A combined row r maps to vision_parts[r // 5000][r % 5000].
-# Process the blocks in order to avoid allocating a full concatenated array.
-```
-
-### Download pooled ImageNet features
-
-```python
-import json
-from pathlib import Path
-import numpy as np
-from huggingface_hub import snapshot_download
-
-root = Path(snapshot_download(
-    repo_id="336labs/VisionEncoder-Features",
-    repo_type="dataset",
     allow_patterns=[
-        "encoders.json", "README.md", "LICENSES.md",
-        "imagenet-200k/features/001_clip_openai__l14.npy",
-        "imagenet-200k/metadata/001_clip_openai__l14.json",
-        "imagenet-200k/labels.npy",
-        "imagenet-200k/source_indices.npy",
-        "imagenet-200k/samples/**",
-        "imagenet-200k/protocols/**",
+        "encoders.json",
+        "lcs-1k/vision/clip_openai__l14*",
+        "lcs-1k/text/qwen3*",
+        "lcs-1k/samples/**",
     ],
-    local_dir="features",
-))
-base = root / "imagenet-200k"
-features = np.load(base / "features/001_clip_openai__l14.npy",
-                   mmap_mode="r", allow_pickle=False)
-labels = np.load(base / "labels.npy", allow_pickle=False)
-protocol = json.loads((base / "protocols/knn_seed42.json").read_text())
-train_rows = np.asarray(protocol["train_indices_by_shot"]["20"])
-query_rows = np.asarray(protocol["query_indices"])
-```
-
-### Load the sample tables with Hugging Face Datasets
-
-```python
-from datasets import load_dataset
-
-lcs_samples = load_dataset("336labs/VisionEncoder-Features",
-                           "lcs-1k", split="train")
-lcs10k_samples = load_dataset("336labs/VisionEncoder-Features",
-                              "lcs-10k", split="train")
-imagenet_samples = load_dataset("336labs/VisionEncoder-Features",
-                                "imagenet-200k", split="train")
+)
+vision = np.load(f"{root}/lcs-1k/vision/clip_openai__l14_patch_n1000_seed42.npy", mmap_mode="r")
+text = np.load(f"{root}/lcs-1k/text/qwen3_penultimate_mean_n1000_seed42.npy", mmap_mode="r")
 ```
 
-The three dataset configurations expose **sample metadata tables** in the Dataset Viewer and `load_dataset`. Download the large feature arrays separately and load them with NumPy. A complete download uses `snapshot_download` with `repo_type="dataset"` and no `allow_patterns`; selective downloads are usually sufficient.
+Use the paths in the encoder catalog below to select other features. The HF Dataset Viewer displays sample metadata; the feature arrays are separate `.npy` files. Raw images are available from the linked source datasets.
 
 ## Encoder Catalog
 
-`encoders.json` maps all 70 tokenizer IDs to their LCS-1000, LCS-10000 block, and ImageNet array paths, metadata, shapes, and upstream encoder sources. The `lcs_patch_n10000.blocks` entries specify each block's global row range and shared sample manifest. Upstream links identify the encoder projects and downloadable weights; an exact historical weight revision is provided only where the original audit recorded one.
+[encoders.json](encoders.json) lists feature paths, shapes, metadata, and upstream encoder sources for all 70 encoders.
+
+<details>
+<summary>Browse all encoders and feature download links</summary>
 
 | Tokenizer ID | Family | LCS `[T, D]` | ImageNet `D` | Feature files |
 | --- | --- | --- | ---: | --- |
@@ -393,26 +183,83 @@ The three dataset configurations expose **sample metadata tables** in the Datase
 | [`webssl_mae300m_full2b_224`](https://huggingface.co/facebook/webssl-mae300m-full2b-224) | SSL | `[196, 1024]` | 1024 | [patches](lcs-1k/vision/webssl_mae300m_full2b_224_patch_n1000_seed42.npy) · [pooled](imagenet-200k/features/069_webssl_mae300m_full2b_224.npy) · [10k blocks](https://huggingface.co/datasets/336labs/VisionEncoder-Features/tree/main/lcs-10k/vision/webssl_mae300m_full2b_224) |
 | [`webssl_mae3b_full2b_224`](https://huggingface.co/facebook/webssl-mae3b-full2b-224) | SSL | `[256, 3072]` | 3072 | [patches](lcs-1k/vision/webssl_mae3b_full2b_224_patch_n1000_seed42.npy) · [pooled](imagenet-200k/features/070_webssl_mae3b_full2b_224.npy) · [10k blocks](https://huggingface.co/datasets/336labs/VisionEncoder-Features/tree/main/lcs-10k/vision/webssl_mae3b_full2b_224) |
 
-## Dataset Creation and Provenance
+</details>
+
+<a id="datasets-required-by-the-evaluation-pipelines"></a>
+
+## Data Sources
+
+<details>
+<summary>MLLM training, evaluation benchmarks, and baseline datasets</summary>
+
+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 method
+
+| 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](https://www.image-net.org/download.php); the released 200k cache contains training images only |
+| Alignment probing | COCO Karpathy captions; DreamLIP CC3M for the cross-dataset variants | [COCO / Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [DreamLIP captions](https://huggingface.co/datasets/qidouxiong619/dreamlip_long_captions) |
+| AC Policy / Law of Vision Representation | LCS-558K, Stage-1 projectors, and SPair-71k | [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain), [Model Zoo](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo), and [SPair-71k](https://cvlab.postech.ac.kr/research/SPair-71k/index.html) |
+| Mid-training loss | LCS-558K images/captions and projector-training logs | [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain); training instructions in the [code repository](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#running) |
+| TokBench | Original images/annotations and matching tokenizer reconstructions | [TokBench](https://huggingface.co/datasets/Junfeng5/TokBench) |
+| MLLM training / evaluation | Training: LCS-558K and filtered mix665k; evaluation: the 11 benchmarks below | [Model Zoo](https://huggingface.co/336labs/VisionEncoder-to-MLLM-ModelZoo) and the linked dataset sources |
+
+ImageNet kNN features have class labels but no paired captions, so they cannot supply an alignment probe's image–text inputs on their own. A full linear-probe benchmark also needs independent validation data.
+
+### Raw data folders
+
+The following are example local folders. Keep the image subdirectories recorded in each dataset's annotations.
+
+```text
+data/
+  instructions/pretrain/blip_laion_cc_sbu_558k.json
+  instructions/finetune/llava_v1_5_mix665k_drop_ge8kchars.json
+  instructions/test/<dataset>.tsv
+  images/pretrain/
+  images/finetune/
+  images/test/
+  imagenet/{train,val}/
+  coco/
+  cc3m/
+  SPair-71k/
+  tokbench/
+```
+
+**MLLM pretraining:** obtain the JSON and `images.zip` from [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain). Image names in the JSON are relative to `data/images/pretrain/`.
 
-This release assembles the project's existing feature caches. Feature arrays are published without additional pooling, quantization, or re-extraction. Released metadata uses repository-relative paths for shared samples and arrays. Inaccessible historical local paths are marked `source-local:<filename>`.
+**MLLM finetuning:** obtain [mix665k](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_v1_5_mix665k.json) and its COCO, GQA, OCR-VQA, TextVQA, and Visual Genome images using the [LLaVA data preparation instructions](https://github.com/haotian-liu/LLaVA#visual-instruction-tuning). The project uses `llava_v1_5_mix665k_drop_ge8kchars.json`, which excludes 395 examples with at least 8,000 conversation-text characters. Image names in this JSON are relative to `data/images/finetune/`.
 
-SHA-256 hashes of original metadata and original sample manifests are retained separately from the hashes of sanitized release files. For text features, `extraction_identity` preserves the original extraction identity; its manifest hash refers to the original manifest. The top-level `sample_manifest_sha256` refers to the sanitized published manifest, and `source_sample_manifest_sha256` retains the original hash.
+### Downstream MLLM benchmarks
 
-All arrays passed finite-value checks before upload. Shapes/dtypes and row-order fingerprints were checked; the 70 encoder IDs match across both subsets. The ImageNet subset has exactly 200 examples per class; the kNN pool/query split is disjoint and shot subsets are nested. The repaired UniAR array matches its recorded repaired hash.
+Place the prepared benchmark TSVs in `data/instructions/test/`. Obtain the matching data from these sources:
 
-The added 140 patch arrays and six last-token text arrays passed full-file finite-value and SHA-256 checks before publication. All 70 encoder pairs have matching block shapes and dtypes. The component manifests concatenate exactly to the 10,000-row manifest, with no overlap in source index, image ID, or text ID. Text file hashes match their original extraction audits. Full source-index arrays are derived from the ordered canonical manifests, preserving the original caches.
+| Benchmark / required TSV | Source |
+|---|---|
+| `MMMU_TEST.tsv` | [MMMU](https://huggingface.co/datasets/MMMU/MMMU), test split |
+| `MMBench_TEST_EN_V11.tsv` | [MMBench](https://github.com/open-compass/MMBench), English test v1.1 |
+| `VQAv2_VAL.tsv` | [VQA v2](https://visualqa.org/download.html), validation split |
+| `ScienceQA_VAL.tsv` | [ScienceQA](https://github.com/lupantech/ScienceQA), validation split |
+| `ChartQA_TEST.tsv` | [ChartQA](https://github.com/vis-nlp/ChartQA), test split |
+| `DocVQA_VAL.tsv` | [DocVQA](https://www.docvqa.org/datasets/docvqa), validation split |
+| `TextVQA_VAL.tsv` | [TextVQA](https://textvqa.org/), validation split |
+| `POPE.tsv` | [POPE](https://github.com/RUCAIBox/POPE) |
+| `GQA_TestDev_Balanced.tsv` | [GQA](https://cs.stanford.edu/people/dorarad/gqa/download.html), balanced test-dev split |
+| `MSCOCO_KARPATHY_TEST.tsv` | [Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [COCO images](https://cocodataset.org/#download) |
+| `FLICKR30K_KARPATHY_TEST.tsv` | [Karpathy splits](https://cs.stanford.edu/people/karpathy/deepimagesent/) and [Flickr30k images](https://shannon.cs.illinois.edu/DenotationGraph/) |
+
+Prepare images referenced by the TSVs under `data/images/test/`. Folder names follow the benchmark loader: for example, MMBench v1.1 uses `MMBench_V11/`. TSVs with embedded images can be decoded by the loader. GQA also accepts the filename `GQA_TESTDEV_BALANCED.tsv`.
+
+Use the [VLMEvalKit dataset loaders](https://github.com/open-compass/VLMEvalKit/tree/main/vlmeval/dataset) for benchmark preparation. Original dataset downloads must be converted to the required TSV format; changing a filename alone is insufficient. Configuration and experiment commands are in the code repository's [Setup](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#setup) and [Running](https://github.com/JuntaoTang/MLLM-VisionEncoder-Eval#running) sections.
 
-The dataset inventory records array shapes, dtypes, byte sizes, SHA-256 hashes, and nonfinite-value counts. `checksums.sha256` verifies all listed files; it excludes itself and `FILES.json` to avoid circular hashes. Sources were preserved during publication.
+</details>
 
-## Intended Uses and Limitations
+## Metadata and Integrity
 
-- **RAVEL and alignment probes:** use paired LCS visual and text features, preserving the shared row order. LCS-1000 uses mean-token text; the historical LCS-10000 cache uses last-token text. Keep these surfaces distinct. Captions and text embeddings are required for cross-modal probes.
-- **kNN and linear probes:** use ImageNet pooled features with shared labels and the provided split. A full ImageNet linear-probe benchmark additionally requires the appropriate training and validation data.
-- **Historical patch provenance:** the original visual audits do not record a uniform feature-layer specification. Missing extraction-layer, processor, or weight-revision fields remain explicitly unverified in this release. The arrays should be interpreted using available per-encoder metadata and the extraction code, rather than assuming a common layer/readout.
-- **UniAR correction (LCS-1000 only):** the published LCS-1000 UniAR array uses the documented legacy BSQ deepstack concatenation. One NaN at `[630, 295, 1051]` was replaced by `1.7109375`, obtained by re-extraction with the official model. Its metadata retains original/repaired hashes, pinned model/source revisions, weight hash, and verification of the remaining 1,151 components in that BSQ vector.
-- **Coverage and bias:** LCS-1000 and LCS-10000 are captioned-image samples; ImageNet200k is a class-balanced training subset. Findings depend on these data sources and selected encoders. Features inherit the content and representation biases of the upstream datasets and models.
-- **Scope:** this release contains the three feature subsets described above. Obtain raw images and encoder weights from the upstream sources, and MLLM checkpoints from the Model Zoo. Derived PCA/distance caches and additional experiments beyond the documented subsets are outside the current release.
+The released arrays retain their original storage precision. Per-file metadata and sample manifests record available extraction and sampling information. [FILES.json](FILES.json) and [checksums.sha256](checksums.sha256) provide file sizes and SHA-256 checksums.
 
 ## Licensing and Attribution
 

checksums.sha256

@@ -1,6 +1,6 @@
 d9d3820e2ad8c5692d2d4f33db110efeafe31a97b1172de592ad9743b6e8d1e4  .gitattributes
 da50ab668f491868ce523b2d62fdd65de12e047163150a11f2bd5ba05857ed38  LICENSES.md
-f157bc76e1cb382bb005a48bc637e81c5bf2a2af0ea9f2f7d0b3a50d2092e9fa  README.md
+1c1b8143eb2af87e8989b952879eea7648125b85e59284793e5dcff79f2f072d  README.md
 46e6a9a7cc6cde8ed94431c91dce1b389e7477a03de2f6626efa9a5d80833e8e  encoders.json
 55cd1a3a1469b858ac539851c34d2f88b5e2cda5737998154a658120820ddcd9  imagenet-200k/export_manifest.json
 c0f64e28098e460ee5cd5716fefd2b1577c54755c5c3ea80dda9a843182ec9fd  imagenet-200k/features/001_clip_openai__l14.npy