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.
Updated 2026-10-18 · How the library is built
3,164 datasets, sorted by most downloaded.
This public dataset repository serves as the persistent cloud memory for the Master Brain Autonomous Orchestrator. • Repository: nagarhimanshu37/brain-memory • Total Stored Memories: 1 • Last Synchronized: 2026-09-27 09:11:34 UTC
The estate's prod-config knowledge graph plus its canonical textual-address layer: a knowledge graph over the estate's own datasets/tools/routes, a deterministic verse/token/pericope address resolver, and a set of query-time registries. Part of the NuBerea curated corpus estate of biblical and historical texts. The full specification (per-config documentation, schemas, methodology, and upstream provenance) is available in the gated SPEC accompanying this dataset. Estate-authored KG projection, address layer, and registry structure: CC BY 4.0. This repo ships ids, ordinals, hashes, and short (1-6 word) pericope/chapter/book section titles — no verbatim source text. Upstream editions that…
An English-language pretraining corpus prepared for controlled experiments with approximately 125M-parameter GPT-2 models based on karpathy/nanoGPT. The name refers to the approximately 5B-token mixture selected before final BPE tokenization. With the included 12,288-token BPE tokenizer, the packaged nanoGPT training split contains 5,396,605,407 tokens. The dataset provides both reusable source text and ready-to-train nanoGPT binaries. - documentid: string - sourceid: string - text: largestring - 12,029 merges - 256 byte tokens - 3 reserved special tokens -: ID 12,285 -: ID 12,286 -: ID 12,287 - One separator appended after every packaged document The tokenizer was trained on a…
This dataset contains Direct Preference Optimization (DPO) pairs translated and culturally adapted into Swedish from the original argilla/ultrafeedback-binarized-preferences-cleaned dataset. The dataset contains the following fields: - prompt: The culturally adapted Swedish prompt. - chosen: The high-quality, authentic Swedish response. - rejected: The poorly translated, uncalibrated response. - language: The target language of the row. This dataset inherits the license from its source dataset argilla/ultrafeedback-binarized-preferences-cleaned (MIT).
Experiment logs, synced 2026-10-17 20:46:24 every few minutes. tdprmresults.tar (5946 MB, tar -xf tdprmresults.tar -> tdprmresults/) holds every metric, trace, evaluation, state, queue and service log of the run; no model weights, environments or dataset text. Updated 2026-10-17 20:46:22 · refreshes automatically Training queue (finished): training runs (97) done 97; other jobs (data, states, base evals) done 44 Step-credit evaluation queue (finished): 10 jobs, done 10 (see Step-credit evaluation) Greedy accuracy, v1 and v2 methods in separate tables. Δ: v1 methods vs grpo, v2 methods vs grpov2 (same cell), GRPO rows vs the base model. pass@8 on the in-domain set; mean@8 / pass@8 for every…
WebsiteBench Harbor trajectories The historical section contains one final valid GLM 5.3 Flash result per task. Every directory follows task-job/glm-5.3-flash/run1/ and contains the Harbor job archive, the matching original English cc-analysis.md, its metadata, and backup provenance. Archives expand under exactly the same directory structure. run1 is the user-requested publication number, not a claim that a recovered trajectory was originally generated as a fresh first attempt.
ALIEN — data and results Data, model adapters and experiment outputs for ALIEN: Aligned Entropy Head for Improving Uncertainty Estimation of LLMs (submitted results + revision experiments). Code: the ALIEN repository, revision/ folder. Layout folder content results/ Published runs used for the paper tables: resultsonembeddings/grid (ALIEN, its variants and ablations; bestrun.pickle = test scores + metricdf with the selected hyper-parameters)
Public, versioned machine artifacts for the Sharpa functional-reorientation pipeline. Source code and download validators live in The repository now includes eight Sharpa Stage3 AutoDR residual-teacher checkpoint candidates for generating physics-executed recovery trajectories. Give an Astra agent the following guide as its execution handoff: This release contains the teachers and collection contract. It does not claim that the proposed augmented rollout dataset has already been collected. Artifacts are organized by object and immutable release: - Eight per-phone Stage3 residual-teacher checkpoint candidates: phone02, phone03, phone05, phone07, phone10, phone14, phone17, and phone18. - Each…
A versioned negative-image resource for fire/smoke detector training and false-positive analysis. The current publication contains 12,653 native source-labelled negatives and 195 retained JPEG candidates, in two separate configurations. The repository name does not establish that every image is a very hard negative. The native subset has no individual visual review or difficulty rating. Existing source classification annotations were the admission basis and were not changed into fresh manual annotations. These identifiers are the source namespaces recorded in the retained provenance, not newly inferred labels or authorizations. See source records and the For the separate JPEG collection…
English · 简体中文 Expert demonstrations for robotic value models, covering real-world and simulated manipulation tasks. Each domain has two demonstration-access settings: Full-Shot for training or fine-tuning, and One-Shot for a single task-specific reference or adaptation example. The common release format is LeRobot Dataset v3.0: per-step Parquet data, camera videos, and task/episode metadata. Robot platforms, cameras, and state/action schemas are specified separately for each subset. Available now: Real-World Full-Shot — 2,000 episodes across 20 tasks; Simulation Full-Shot — 1,500 episodes across 15 tasks. Simulation One-Shot is available: 15 episodes / 8,080 rows. Real-World One-Shot…
This repository contains training-video shards prepared for VideoVAE experiments. - Videos available on the packaging host: 27,351. - Each shard contains videos/partN/.mp4 and an internal manifest.jsonl. - Global manifests are under OpenViD-HD/metadata/manifests/videos-min132/. The 81--131-frame expansion is intentionally not included in this release because the complete quality-metric table needed to apply the same gates was unavailable when this subset was packaged. See OpenViD-HD/README.md for the organized videos, VAE latents, and current >=121-frame qualified selection. The older description above documents the original release. See UltraVideo/README.md for the metadata-first…
Two human egocentric stereo episodes recorded on real job sites in Hangzhou, China, released as a public, verifiable sample of a production egocentric data-collection pipeline. This is a sample, not a benchmark. It exists so that anyone evaluating egocentric data vendors can inspect real output — actual frame timing, actual synchronisation, actual annotation depth — instead of reading a capability deck. Independently validated. The dataset passes The same two episodes previously passed an external buyer's incoming data QC (August 2026). Total: 4,914 frames / 163.47 s, two synchronised streams per episode. Two numbers worth pausing on, because they are where most egocentric data fails: - p99…
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.