Claude Skills by feiyang-k

> Variant of `BENCHMARK.md`. The agent is given exactly one of `protocols/plain.md`, `protocols/instruction.md`, or `protocols/skill.md` at launch, depending on the chosen launch profile. The skill-grounded variant additionally exposes a library of paper-derived skill cards mounted at `/workspace/skills/<category>/<paper-slug>/SKILL.md` inside the container — agents are required to read a quota of them before editing curation code. The host-side source of those cards is the repository's top-l...

Use this skill when you need a VQA dataset requiring world knowledge and commonsense reasoning beyond what is visible in the image. Avoid it when your VQA task only requires visual perception without external knowledge.

Use this skill when you need a large-scale video activity understanding dataset with temporal annotations for training video-language models. Avoid it when short-clip action recognition is sufficient.

Use this skill when you want to use Whisper ASR to transcribe speech from videos for creating video-text training data. Avoid it when your videos do not have speech or you have existing transcripts.

Use this skill when you want to build a biomedical CLIP model trained on 15M scientific image-text pairs from PMC. Avoid it when you need a general-purpose CLIP model or lack access to PMC data.

Use this skill when you need a benchmark revealing perception gaps in VLMs — tasks that are trivially easy for humans but hard for models. Avoid it when standard VLM benchmarks show sufficient performance.

Use this skill when you need a vision-centric evaluation benchmark that tests visual perception rather than language ability in VLMs. Avoid it when standard VLM benchmarks are sufficient.

Use this skill when you want to scale up an image-text dataset from 3M to 12M by relaxing filtering criteria to capture long-tail visual concepts. Avoid it when you need very clean data and cannot tolerate increased noise from relaxed filtering.

Use this skill when you need a QA dataset about charts requiring both visual perception and logical reasoning. Avoid it when you need VQA for natural images rather than charts and graphs.

Use this skill when you need a standardized benchmark suite for evaluating CLIP-style vision-language encoders across dozens of datasets. Avoid it when you only evaluate on ImageNet zero-shot.

Use this skill when you need gold-standard image captions collected through crowdsourcing for VLM training and evaluation. Avoid it when web-crawled captions are sufficient or you have your own caption data.

Use this skill when you need a foundational multi-purpose vision dataset with captions, detections, segmentation, and keypoints for VLM development. Avoid it when you need web-scale data rather than a carefully annotated 330K image dataset.

Use this skill when you want to build an image captioning dataset from web alt-text using automated cleaning and hypernym replacement. Avoid it when you need fine-grained or domain-specific captions that web alt-text cannot provide.

Use this skill when you need a large-scale open image-text dataset from Common Crawl as an alternative to LAION with transparent collection methodology. Avoid it when LAION-5B or other existing datasets are sufficient for your needs.

Use this skill when you need a VQA dataset on document images for training and evaluating document understanding VLMs. Avoid it when you need scene-level VQA rather than document understanding.

Use this skill when you need a massive egocentric video dataset with diverse annotations for training video-language models. Avoid it when you do not work with egocentric/first-person video or need third-person video data.

Use this skill when you want to curate diverse interleaved multimodal data combining web pages, image-text pairs, and video for generative pretraining. Avoid it when you only need paired image-text data without interleaved document structure.

Use this skill when you need to collect and train on interleaved image-text web data for few-shot visual learning. Avoid it when you only have paired image-caption data without interleaved document structure.

Use this skill when you need a dataset linking noun phrases in captions to bounding box regions in images for visual grounding. Avoid it when you do not need phrase-level grounding or the 31K image scale is too small.

Use this skill when you need to efficiently download billions of images from URLs for building large-scale image-text datasets. Avoid it when you already have images locally or only need a small number.

Use this skill when you need a benchmark testing VLM robustness across multiple visual tasks with distribution shifts. Avoid it when standard in-distribution benchmarks are sufficient.

Use this skill when you need a diagnostic benchmark testing VLM vulnerability to both language hallucination and visual illusion. Avoid it when you only need standard hallucination evaluation (use POPE instead).

Use this skill when you want to mine narrated instructional videos from YouTube as video-text training data at hundred-million scale. Avoid it when you need clean video captions rather than noisy ASR transcripts.

Use this skill when you need a large-scale hierarchical image dataset organized by WordNet synsets for vision model pretraining and evaluation. Avoid it when you need image-text pairs rather than image-label data.

Use this skill when you need a 234M video-text dataset constructed with multi-scale captioning from ASR, metadata, and generated descriptions. Avoid it when you only need image-text data or a small video dataset.

Use this skill when you need a large-scale action recognition dataset with 400 categories and 300K+ video clips for video encoder pretraining. Avoid it when temporal reasoning rather than action recognition is your focus.

Use this skill when you need to build a multi-billion scale open image-text dataset from Common Crawl using CLIP score filtering. Avoid it when you need a small curated dataset or cannot handle the infrastructure for billions of image downloads.

Use this skill when you want to align multiple modalities (video, audio, depth, thermal, infrared) to language for unified multi-modal understanding. Avoid it when you only need image-language alignment without other modalities.

Use this skill when you need to construct a large-scale image-text dataset from the web for contrastive vision-language pretraining. Avoid it when you already have curated paired data or need fine-grained region-level annotations.

Use this skill when you need a unified evaluation framework for consistently evaluating VLMs across dozens of benchmarks. Avoid it when you only evaluate on 1-2 benchmarks.

Use this skill when you need a large-vocabulary instance segmentation dataset with 1,200+ categories handling the long tail of visual concepts. Avoid it when COCO's 80 categories are sufficient for your needs.

Use this skill when you need a benchmark testing mathematical reasoning in visual contexts including charts, plots, diagrams, and geometry. Avoid it when you do not need mathematical reasoning evaluation.

Use this skill when you need a large-scale medical image-report dataset for training biomedical VLMs on chest X-ray understanding. Avoid it when you do not work with medical imaging.

Use this skill when you need a benchmark evaluating VLMs on tasks requiring integrated use of multiple capabilities simultaneously. Avoid it when individual capability testing is sufficient.

Use this skill when you need a multi-dimensional benchmark evaluating VLMs across 20+ ability dimensions with robust circular evaluation. Avoid it when you only need a single-dimension evaluation or have simpler benchmarks.

Use this skill when you need a comprehensive VLM benchmark testing both perception and cognition abilities with yes/no questions. Avoid it when you need open-ended evaluation or have specific benchmark needs.

Use this skill when you need a benchmark testing expert-level multimodal understanding across 30+ subjects using college exam questions. Avoid it when you need basic visual QA evaluation rather than expert-level assessment.

Use this skill when you want fully open-source VLM training data (PixMo) that powers a state-of-the-art model without proprietary data dependencies. Avoid it when you are fine using proprietary data sources.

Use this skill when you need a video description dataset with 10K web clips and 200K sentences for training and evaluating video-language models. Avoid it when you have sufficient video captioning data.

Use this skill when you want to augment C4 text corpus with interleaved images placed at relevant positions for multimodal pretraining. Avoid it when you only need paired image-text data or cannot process the billion-scale corpus.

Use this skill when you need a visual reasoning benchmark where the model must determine if a statement is true for a pair of images. Avoid it when single-image reasoning evaluation is sufficient.

Use this skill when you want a dataset of 10K personally written instruction-response pairs created entirely by skilled humans, not AI. Avoid it when AI-generated instruction data is acceptable for your needs.

Use this skill when you need a captioning benchmark testing generalization to novel objects not seen during training. Avoid it when you only evaluate on in-domain captioning.

Use this skill when you need a comprehensive multimodal autonomous driving dataset with cameras, LiDAR, radar, and 3D annotations for driving VLMs. Avoid it when driving data is not relevant to your work.

Use this skill when you want to build an open web-scale dataset of interleaved image-text documents for multimodal pretraining. Avoid it when you only need paired image-caption data without document context.

Use this skill when you need a large-scale detection dataset with 365 categories for training open-vocabulary detection models used in VLM pipelines. Avoid it when you only need the 80 COCO categories or a smaller detection dataset.

Use this skill when you need a VQA dataset about book covers requiring OCR to read titles, authors, and other text. Avoid it when you need general VQA or document VQA rather than book cover understanding.

Use this skill when you need a VQA dataset where questions require external knowledge sources like Wikipedia to answer. Avoid it when your VQA task does not require external knowledge retrieval.

Use this skill when you want to aggregate robot learning data across different embodiments and labs for training generalist robot policies. Avoid it when you are not working with robotics or have data from a single robot type.

Use this skill when you need a massive multi-label detection dataset with 9M images, 600 categories, and visual relationship annotations. Avoid it when COCO or Objects365 provide sufficient detection coverage.