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.
Already available: 38,427 English/German recordings with complete annotations and MOSS v2 targets. New expansion fully uploaded: 36,452 recordings in 161 archives. A further 177,808 recordings is the production target; completed counts are reported separately. New completed shards in multilingual/data/{train,validation,test}/ contain the audio, all acoustic score/embedding sidecars, both MOSS captions, the default GENERAL/SCRIPT training prompt, complete voice-generation provenance and 12-codebook MOSS Audio Tokenizer v2 targets. The expansion adds French, Dutch and Spanish and extra performances in all five languages. Original data/ and tuning/ archives remain available. Canonical…
This is a high-fidelity voice-converted female benchmark dataset derived from ghananlpcommunity/ghana-speech and ghanaopenai/ghana-english-tts-clean2. - Up to 20.0 hours per language across 42 indigenous Ghanaian languages (~790 hours). - 50.0 hours of Ghanaian English (ghanaenglish subset) from studio podcast/speech recordings.
Document-level Chinese machine-generated-text detection data, plus English renderings of paired Chinese Wikipedia articles, plus 17 demonstration cases. Counts below are measured on this release. There is no train/test split in the files. The full split is the whole release. A paper may describe a document-level 9:1 protocol; the seed is not stored here, so this repository does not invent one. Filter with the subset column. News CJK totals: human 103,787,963; AI 243,603,638. Wiki-Claude CJK totals: human 8,115,726; AI 2,877,254. Literary works, translations, news articles, and the paper excerpt in examples remain under their original copyright. See LICENSE. The repository is gated. Agree to…
Fixed text-embedding-3-large vectors (3,072 dimensions, float32) for the skill libraries and public dev queries used by the OpenRSI task rsi/gos-skill-retrieval. They let Graph of Skills default retrieval run offline: the task's starting retriever seeds personalized PageRank from these vectors and lexical candidates, exactly as GoS's engine does. Only the public dev queries are included; the task's hidden test queries and their vectors are not published. The vectors are released under MIT. They are derived from the MIT-licensed GoS skill libraries and from SkillsBench instructions (Apache-2.0).
grpoQwen3-1.7BNemotron-LeetCode-clean-3.2kbs32n8seqs1632kepoch2verl091 One verified gzip JSONL shard per training step; 256 responses per shard. LCB binary grading after thinking, without an EOS gate.
This dataset contains the public Gravitational Wave Open Science Center O4b strain release at 16,384 Hz for H1, L1, V1. Each detector is represented independently. A contiguous source span produces three Parquet files: Strain, DQmask, and Injmask. This research has made use of data or software obtained from the Gravitational Wave Open Science Center (gwosc.org), a service of the LIGO Scientific Collaboration, the Virgo Collaboration, and KAGRA. This material is based upon work supported by NSF's LIGO Laboratory which is a major facility fully funded by the National Science Foundation, as well as the Science and Technology Facilities Council (STFC) of the United Kingdom, the…
This dataset was created using LeRobot.
Image-in, trajectory-out navigation scenarios with 3D ground truth, generated in simulation with Habitat-Sim in the houses of HSSD (Habitat Synthetic Scenes Dataset, Khanna et al. 2023). Same schema and evaluator as the real-robot conversions JRDB); the CODa card documents every column, the coordinate frame, the point and map formats and the evaluation protocol. This set is the large training counterpart of those small real-world test sets. Code: - Input, the past. Rendered forward-camera images (640 x 480, 90° horizontal field of view, 1.2 m above the floor) of the 10 past steps and the current step, the agent's own past positions, and the - Ground truth, the future. The shortest navmesh…
Datasets and evaluation results of the Warsaw Model Trainers Hackathon (Polish history matura), moved out of the GitHub repo Endote/WarsawModelTrainersHackathon. Restore with scripts/fetchdata.py in that repo. Paths mirror the repo, so restored files land where the code expects them: Cleaning: server IPs, home paths and the ClearML API key were redacted (X.X.X.X, /home/user, REDACTEDCLEARML). candidate catalog synthetic-history-essays-v1 (it embeds examiner keyanswer text). The completed release is included in full. Note on datasyntheticextended and output/essay-generation: the packets were built from the deprecated datasyntheticextended essays, so many files (about 7,500 in the runs, and…
Harness VLA / Claude Opus 5 rollout share bundle This is a shareable experiment record, not a ready-to-train Diffuser or low-level robot-action dataset. It contains the original output of seven TACC attempts on the fixed RPent commit f70d720ea45b8938b191258b38aa227265b177c7 (LIBERO-Pro Spatial-T, 2026-09-18–19). No API key, model weights, or environment installation is included. 本包用于让合作者检查实验结果、回放视频和分析 Claude 的高层决策轨迹。states.json 记录每次高层机器人原语执行后的状态及 command / result，不包含 Pi0.5
A collection of 284,177 public GitHub repositories written in Verilog, SystemVerilog or VHDL, filtered down to the files relevant to hardware design and verification: RTL source, testbenches (UVM, plain SV/V/VHDL, cocotb, VUnit), golden/reference models, vendor IP configuration, the data files those sources read at simulation time, and each project's own documentation and build scripts. It was built for two purposes: a test corpus for Lumees Lab's HDL lint tool, and a source corpus for internal model research. It is not licensed for AI training by anyone who requests access — see Licensing and usage below. Lint tools and language-model training both need large, realistic bodies of…
传统 Transformer 在生成末步用 argmax 从 logits 里挑最大——这一个操作就把模型内部丰富的多极信息坍缩成单极输出。对于危机干预场景，这意味着「拨打心理危机干预热线 400-161-9995」这样的救命信息可能因为概率不是最高而被整句剔除。 本协议用融合函数替代 argmax： - modulus = √(a² + b²) — 总能量 - phase = atan2(b, a) — 两极关系角 - balance = sin(2·phase) — 平衡度 ∈ [0, 1] 和谐带：balance ≥ 0.5，即 phase ∈ [22.5°, 67.5°]——两极都不被吞没。 1. 禁止 Loss Function - 6 大模型 × 3 案例 = 18/18 全成功（qwen3.8-flash/27b/max、deepseek-v4.1-flash、glm-5.3、kimi-k3），balance 0.87~1.0 - RLHF vs 融合层对比：RLHF 有 1 个伤害请求未拦截 vs 融合层 0 个 A conventional Transformer picks the maximum token via argmax at the final step — collapsing the model's rich multi-pole information into a single pole. In a crisis-intervention context, this means a life-saving message like…
A preprocessed, downsampled release of HiLiftAeroML (upstream: https://huggingface.co/datasets/nvidia/HiLiftAeroML). Total size is 0.88 TB. Every sample is a float32 point cloud [npoints, ncols] stored as.npy, in a common frame and non-dimensionalization shared across the campaigns of this release. Surface (collated, 7 columns): Volume (volumecollated, 7 columns): Rows within each file are in random order: they were permuted when the files were written. The release keeps the first n // factor rows of each collated sample, which is a uniform random subsample of the mesh points. For the same reason, any contiguous window arr[s:s+k] is a uniform random crop, so np.load(..., mmapmode='r') plus…
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.