Category

OCR-Synthetic-Multilingual-v1

Provides 12.26M synthetically generated multilingual OCR samples (en/ja/ko/ru/zh) with word/line/paragraph bounding boxes and reading-order graphs, packaged as HDF5 shards for training detection, recognition, and layout models; licensed CC BY 4.0.

ParseBench

Benchmarks document-parsing systems on real-world enterprise PDFs and images—evaluates tables, charts, content faithfulness, semantic formatting, and visual grounding with human-verified, rule-level tests. Ships with ~2,000 pages, ~169K test rules, and an open evaluation framework for end-to-end pipeline scoring.

Zero-to-CAD 100K

Curated 100K subset of geometrically diverse CAD construction sequences sampled from a 1M agentically synthesized corpus — each item includes executable CadQuery scripts, 8 rendered views, STL/STEP exports, and precomputed DINOv3 embeddings for retrieval and benchmarking.

Jagle: Building a Large-Scale Japanese Multimodal Post-Training Dataset for Vision–Language Models

Provides a ~9.2M-instance Japanese multimodal post-training dataset for vision–language models, combining image–text pairs, PDF corpora and generated VQA to improve Japanese VLM performance; access is restricted by Japanese copyright (download via llm-jp GitLab).

Zero-to-CAD 1M

Provides one million executable, human-readable CadQuery construction sequences synthesized by an LLM-in-the-loop—each sample includes renders, STL/STEP exports, precomputed DINOv3 embeddings and a FAISS index. Designed for training and benchmarking text/image→3D and CAD-program generation models (Apache-2.0).

GLM-5.1-1000000x

Provides 1,003,589 full chain-of-thought reasoning traces and final answers generated by GLM-5.1, split into main/Math/PHD-Science/Multilingual-STEM subsets. Useful for instruction-tuning, supervised fine-tuning, and reasoning experiments; released under Apache-2.0.

Türkçe Atlas — Instruct SFT

Provides 336,146 Turkish instruction-following chat examples (system→user→assistant) for supervised fine-tuning; single train split (no validation/test), reported MIT license, diverse tasks (rewrites, summarization, QA) and a uniform system prompt that may bias model behavior.

RSRCC

Provides paired before/after satellite images with question–answer annotations for semantic change understanding. Includes Yes/No and multiple-choice formats, delivered in Hugging Face datasets (streaming-friendly), suited for remote-sensing multimodal VQA and semantic change captioning research.

Usenet Corpus 1980–2013

Provides deduplicated, sanitized Usenet posts (1980–2013) for language-model pretraining and linguistic research. Includes a ~103.1B-token full corpus (408M posts) with freely downloadable sample files; full corpus access requires a license and PII redaction was applied.

SWE-ZERO 12M Trajectories

Provides ~12.29M execution‑free agentic coding trajectories (≈112B tokens) sampled from 122K GitHub PRs to mid‑train code and agent models. Uses bash-only actions (grep, git, sed, etc.) so it scales without Docker; trajectories are unverified and intended for mid-training rather than final SFT.

Open-SWE-Traces

Provides 207k+ LLM-generated agent trajectories of code edits and tool interactions for training and evaluating software-engineering agents. Collected via OpenHands and SWE-agent using Qwen3.5-122B and MiniMax-M2.5, multilingual across nine languages and released under CC BY 4.0.

GLM-5.1-Reasoning-1M-Cleaned

Provides a cleaned, SFT-ready collection of ~746k GLM-5.1 reasoning traces for instruction tuning and reasoning distillation. Normalizes varied chain-of-thought formats into a single conversations/input/output schema and preserves four focused subsets (main, PHD-Science, Multilingual‑STEM, Math).