AI Training Datasets

Dataset

koskari

Koskari

Verified Koskari language examples exported from the Koskari conformance corpus. Each row traces to a corpus record with machine-checked expectations at one or more language phases (read, expand, eval, error, invalid). Every corpus record. One row per record regardless of which verify stages are present. Missing stages are null. User documentation pages (Markdown) from the source repository's docs/reference/, docs/start/, docs/guide/, docs/patterns/, and docs/agents/ trees. One row per page. The path column matches the strings in the corpus config's references column verbatim, so a record's references resolve directly to their documentation content. The agents section is written for AI…

Teleoperated bimanual manipulation demonstrations collected in MuJoCo with a BimanualYAM robot, packaged in LeRobot v3.0 format for pi0.5 training. Three RGB views per frame — camerafront, cameraleft, cameraright (224x224, H.264). Robot state and action are both 14-dimensional: 6 arm joints plus 1 gripper per arm. A 128-dimensional non-robot state carries object/scene information for diagnostics. Actions are stored as absolute joint targets. The training recipe converts the 12 arm dimensions to deltas relative to the current action-chunk anchor while the 2 gripper dimensions stay absolute. The normalization statistics in meta/stats.json are computed in that 50-step relative space — do not…

Training rollouts of one verl run, one parquet per trainer step (data/step-.parquet), rows in the order the trainer consumed them; the rollout.n samples of a prompt are contiguous and share uid. prompttokenids / responsetokenids are the exact token ids (special tokens kept, no padding) and rolloutlogprobs the sampling-time log-prob of every response token. input / output are decoded with skipspecialtokens=True and are for reading only. score is the raw reward.

Derived caches for two-stage physics-consistent video generation on Physion. stage1flowtargets/ -- analytic ground-truth optical flow reconstructed from the simulator's mesh vertices, quaternion poses and camera matrices by z-buffer rasterisation (Drape uses flex particles + Delaunay). Not RAFT. Each entry holds flow (7,2,160,160) in pixels, forward and cumulative from the first frame at 0.2s..1.4s, plus visibility, objectmask, futuremask, instanceid, centers. Valid only inside objectmask; background is exactly zero. stage2affineconditions/ -- Stage-2 conditions built from the same flow after projecting it onto one 2D affine transform per instance per horizon. Packed as ~1GB tar shards; see…

Dataset

saraiki-dpo-sft-dataset

Muhammad Farjad Ali Raza

Preference data for training and aligning Saraiki language models using Direct Preference Optimization (DPO). Saraiki (سرائیکی) is an Indo-Aryan language spoken primarily in Pakistan, particularly across southern Punjab and neighboring regions. It has a rich linguistic, literary, and cultural heritage, with its own vocabulary, grammatical structures, and regional varieties. Despite its significant speaker community, Saraiki remains a low-resource language in Natural Language Processing (NLP), with relatively limited digital datasets, language models, and computational resources compared with widely supported languages. The Saraiki Direct Preference Optimization Dataset is a curated…

Dataset

PhysMani-Bench

vLAR Group - HK PolyU

Dataset, checkpoints, and demo video for the PhysMani release. Public files intentionally use neutral names. Internal source filenames are not part of the public release interface.

All products and their information from Lidl Switzerland. The entire catalogue and the current special discounts are all included. Check out similar datasets for other grocery stores - https://huggingface.co/datasets/Yelinz/migros-ch-products - https://huggingface.co/datasets/Yelinz/coop-ch-products - https://huggingface.co/datasets/Yelinz/denner-ch-products

Dataset

opened-h200-runs

Kien Vu

Updated 2026-10-07 10:40 UTC. Runs from three Vast boxes, pushed here as they finish: the first H200 box: the FewRel task0 runs and 7 distillation methods, then (when FewRel ends) the TACRED distillation baselines on orders 0-2; a 2x H200 box: the FewRel CL-LoRA methods, then the TACRED CL-LoRA baselines; a 4x H200 box, from 2026-10-06 13:17 UTC: Ours on every dataset (on ACE with its ablations and the memory-0 CL-LoRA rows), and the TACRED

Dataset

gguf-repro-harness

CyberNative AI

We ran the same pinned pipeline twice — two different operators, fresh containers, same commands — on two Apache-2.0 models (Qwen2.5-0.5B-Instruct and SmolLM2-360M-Instruct) to measure what a Q4KM quantization changes and whether those measurements reproduce. What reproduced across both runs (exact, or within ±2%): - the F16 and Q4KM GGUF files are byte-identical (SHA-256); - perplexity on the Wikitext-2 test set (fixed 128 x 512-token budget) is identical to four decimals for each arm between runs; - HellaSwag accnorm over a fixed 200-task subset is identical between runs for each arm (the arms are not expected to match each other; SmolLM2 F16 52.50% vs Q4KM 54.50%); - peak RSS during the…

Dataset · Robotics

DexCanvas

DexGEM Lab

DexCanvas: Bridging Human Demonstrations and Robot Learning for Dexterous Manipulation DexCanvas is a dataset of 111,857 complete right-hand trajectories and 55,442,683 frames of rigid hand–object interaction. Every Parquet row is one trajectory with a shared time axis for hand kinematics, object pose, and per-frame contacts. For each trajectory, the repository also provides the complete dual-camera JPEG stream of the corresponding capture and shape-specific kinematic assets for replay. Camera images are associated at the complete Raw trajectory level. The release does not claim that trajectory frame i corresponds to image i. - 111,857 trajectories; 55,442,683 frames; 96,484,004 contact…

Dataset · Text retrieval

ipfs_kenya_laws_ir

Justicedao

Research retrieval release of endomorphosis/ipfskenyalaws (revision 27b81c642fec385c9123c30b9eb885d093b70f61) packaged as country-laws-ir-graphrag/v1 (layout family skillcenter-huggingface-release/v3 / publicus-ir). Not legal advice. This is a research snapshot. The official gazette / authentic source of Kenya prevails over this corpus. Retrieved documents and graph edges are retrieval evidence only. No legal text was invented. Primary key: entrycid (CIDv1 raw sha2-256 of a canonical identity record). Integer documentindex values are compact shard pointers, not identities. Target Hub id (packaging metadata only): justicedao/ipfskenyalawsir. entrycid, lawcid, recordtype, jurisdiction…

Training rollouts of one verl run, one parquet per trainer step (data/step-.parquet), rows in the order the trainer consumed them; the rollout.n samples of a prompt are contiguous and share uid. prompttokenids / responsetokenids are the exact token ids (special tokens kept, no padding) and rolloutlogprobs the sampling-time log-prob of every response token. input / output are decoded with skipspecialtokens=True and are for reading only. score is the raw reward.

Dataset · Other

x402-settlement-census

Council of AI

We paid 316 conformant x402 hosts as an ordinary buyer and recorded what came back. Not a survey of what hosts advertise — a record of what they did when real USDC arrived. Two in three conformant hosts refused a correctly-signed payment. Being listed in a Bazaar index and answering a valid 402 is not the same as taking money and delivering. Buyer's-eye, exactly as a client experiences it: 1. GET the resource → 402 + a challenge (v2 header, or v1 body) 2. sign an EIP-3009 TransferWithAuthorization for accepts[0], with the EIP-712 domain taken from the challenge itself 3. GET again carrying PAYMENT-SIGNATURE (v2) or X-PAYMENT (v1) 4. record status, bytes, content-type, latency, and the…

Dataset

open-source-scientific-documents

KASYS

This dataset contains approximately 100,000 open scientific PDF documents packaged as a shared retrieval corpus. Train, validation, and test query sets are expected to reference documentid values from this single corpus rather than using separate document splits. - ACL: 20,845 PDFs - Biology: 22,000 PDFs - Engineering: 22,000 PDFs - Medicine: 22,000 PDFs - Physics: 21,667 PDFs The source folders preserve the project corpus identities: ACL computational linguistics papers, arXiv Physics papers, arXiv Engineering papers, PMC OA Biology papers, and PMC OA Medical/Clinical Research papers. The selection and license filtering are performed by the existing project download pipeline before…

Private dataset of LayoutAI physics scene archives exported from E:/layoutairelease. - 1781 scene zip archives total in source; 1144 committed so far. - Remaining 637 files (~55GB) are blocked by the free-tier private-repo storage limit (about 100GB). To finish the upload, either switch the repository to public or upgrade the account storage. - Upload performed with huggingfacehub uploadlargefolder via a local proxy on 2026-09-16.

Dataset · Text retrieval

ipfs_ethiopia_laws_ir

Justicedao

Research retrieval release of endomorphosis/ipfsethiopialaws (revision 8a977a033c19f8f41f551f9d14d1ef1877472412) packaged as country-laws-ir-graphrag/v1 (layout family skillcenter-huggingface-release/v3 / publicus-ir). Not legal advice. This is a research snapshot. The official gazette / authentic source of Ethiopia prevails over this corpus. Retrieved documents and graph edges are retrieval evidence only. No legal text was invented. Primary key: entrycid (CIDv1 raw sha2-256 of a canonical identity record). Integer documentindex values are compact shard pointers, not identities. Target Hub id (packaging metadata only): justicedao/ipfsethiopialawsir. entrycid, lawcid, recordtype…