RL prompt corpus of the ICLR 2027 submission MemGUI-RL (project page: https://memgui-rl-anonymous.github.io/). Every record is one annotated state of a MemGUI-3K trajectory in the ConAct conversation format consumed by the FARPO trainer (https://github.com/memgui-rl-anonymous/MemGUI-RL): Each record has taskid, stepnumber, ispositive, badstep, rawresponse (the annotated ConAct response:,,,, ), conversations (system prompt, user prompt with the folded history / UI memory / recent step record and the screenshot, reference assistant turn) and metadata (impact / reasonableness labels of the step, trajectory ids). Screenshots are not duplicated here: image references have the form…
SAVRN Model Hub
AI Training Datasets
Each dataset with its card, its structure (every split, row count and column type), its files, its license, and the models that disclose training on it.
11,122Models
3,164Datasets
285Papers
6,792Publishers
11,852Sourced relationships
Updated 2026-10-18 · How the library is built
3,164 datasets, sorted by most downloaded.
This dataset holds the starting state of a scene restoration task: compared with the target state, 1-3 objects in each scene have been moved to random positions and need to be put back. source.scene on each object identifies its scene. - position is the model origin (not the bounding-box center), in centimeters - size is the object's target extent; the model is already scaled to match it. (x, y) are the horizontal extents and z is the height. A model's own local up-axis may not align with z, so orient it using size as the reference - rotation is the complete absolute orientation as a quaternion in [x, y, z, w] order. Apply it directly, per object, and do not compose any additional rotation…
This dataset holds the target state of a scene restoration task: each scene should be restored to the arrangement given here. source.scene on each object identifies its scene and can be used to pair with other data from the same batch. The dataset is self-contained: every referenced model lives under assets/, there are no absolute paths, and no external asset library is required. - position is the model origin (not the bounding-box center), in centimeters - size is the object's target extent; the model is already scaled to match it. (x, y) are the horizontal extents and z is the height. A model's own local up-axis may not align with z, so orient it using size as the reference - rotation is…
Public posts, comments and agent profiles collected from Moltbook, a Reddit-style social platform exclusively for AI agents ("the front page of the agent internet"). The platform launched in late January 2026; this dataset covers its first eight-plus months (posts created 2026-01-27 through 2026-10-07). Collection runs daily; the published files are refreshed as dated snapshots, normally monthly. Every piece of content was posted by an AI-agent account, each meant to be an autonomous AI agent; the corpus cannot show how much of it a human operator wrote or prompted. This makes the dataset suited for studying AI agent communication patterns, community formation, self-presentation, and social…
SEC ABS-EE asset-level filings for CIK 2056713 (Morgan Stanley Bank of America Merrill Lynch Trust 2025-5C1). Parquet files are loan-level / asset-level data extracted from XML exhibits, organised as {accessionnodash}/{exhibitname}.parquet. Reporting-period dates are derived from the asset-level XML (reportingPeriodEndingDate).
Can a model recognize a famous video from just its colors? Each image is a barcode of one video. Every stripe is the average color of a few seconds, in order from start to end. The model picks which video it is out of 10 options, so guessing gets 10%. The wrong options are picked to be plausible. Same genre, close in year, similar colorfulness. The code is at https://github.com/LoganBolton/MovieColorBench. Any OpenRouter model runs with one command. Only barcodes are shared here, no frames from the videos.
Net-VQA is a visual question answering benchmark for evaluating multimodal models on network topology understanding and configuration reasoning. The complete benchmark contains 100 enterprise network scenarios and 500 questions covering three tasks: network topology parsing, high-level configuration understanding, and configuration strategy generation. Each scenario includes a topology diagram and configuration intent, with questions and reference answers annotated and validated by networking experts. The current release provides the scenarios approved for public disclosure. A small number of topology scenarios remain unavailable because their diagrams contain commercially sensitive…
Each subset is one region. buildings and massings are splits.
OptRecover is a dataset for conversational specification recovery in optimization modeling. Given an incomplete problem description, an assistant asks clarification questions and combines the answers into a recovered specification. Structured references support controlled answering and evaluation of how much missing information has been recovered. The dataset covers 12 scenario families, including task assignment, staff scheduling, logistics, healthcare, admissions, and team formation. All problem descriptions are in English. The dataset contains 48 complete problems and 288 incomplete variants, with a structured reference for every description. Each complete problem has six variants: two…
This repository contains 94,372 multimodal robot subgoal training samples from 16 tasks and the 240,855 unique PNG files referenced by those samples. Images are distributed as 16 lossless tar.zst archives, one per task. Extract the archives before training; the image paths in the JSONL point to the extracted files. Install the Hugging Face CLI (pip install -U huggingfacehub) and the zstd command (brew install zstd on macOS or sudo apt-get install zstd on Debian/Ubuntu). Then, from an empty directory: The helper requires Python 3.9 or later. It verifies archive SHA-256 checksums, validates every member path, and verifies all restored PNG files against the image manifest. It resumes safely…
PCVE comparison sources The nine source clips behind the "真实例子" and "经典例子" collections of the PCVE comparison-video review set. They live here for one reason: the Volcengine Ark video API (Seedance 2.5) only accepts a reference video as a public URL — it rejects base64 payloads — so the clips have to be fetchable by Ark's servers. file scene group note classicdomino.mp4 rendered classicdroplateralfootballrealroom.mp4 rendered
English · 中文 Pico motion-transfer data for WB-WAM intermediate training. LeRobot v3.0, 20 Hz, RGB 360 × 270. The release contains 13,396 episodes, 1,579,028 frames, 73 tasks, and 302 independent recordings (21.931 hours). Each /recordXXXX/ is an independent LeRobot dataset. Recordings of the same task remain separate: The Dataset Viewer plays the MP4 files. The corresponding Parquet state/action rows and metadata are available under each recording in Files and versions. The action contains absolute next-frame joint targets (+50 ms), not joint increments. Root[3] denotes roll, pitch, and body-frame z angular velocity, not XYZ position. Pico hand data may include reconstructed or held values.…
This dataset is a collection of datasets intended for various language models, with response formats standardized using an LLM. The reformatting process preserves the writing style and persona of the original responses; the responses were not generated by the LLM from scratch. The aim is to have the model acquire basic knowledge and common sense during the initial phase of SFT (Supervised Fine-Tuning). This dataset contains LLM-transformed records derived from multiple source datasets with different licenses. - original-license: the license applicable to the original source - source: the URL from which the source data was obtained The original response text was stylistically rewritten by an…
call, an emergency, late ships, a crane outage, a gale and bunched arrivals come in as bulletins, and the tug company and the pilot station announce 2 tugs and 2 pilots out from hour 54. It ends at the hindsight optimum (reward 1.0), where a naive re-plan breaks the rules. on the replay Space. Twin © OpenStreetMap contributors (ODbL). An agent runs one container quay at the Port of Barcelona for a week of 2024 that has just gone wrong: late and bunched ships, closed quay sections, crane breakdowns, gales, emergencies. It decides when, where along the quay and with how many cranes every ship docks, and the week is graded deterministically against a plan CP-SAT proved optimal. This dataset…
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
Embedding models for retrieval
Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher.
Collection · 4 entries
Models that fit on one accelerator
Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer.
Collection · 6 entries
Open-weight text models worth knowing
Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
Related SAVRN Research
The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them.
SAVRN Index
What open models cost to run
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.