OneKE

OneKE is an open-source knowledge extraction framework built by Ant Group and Zhejiang University and released to the OpenKG community. It converts unstructured documents into structured knowledge across three tasks — named entity recognition (NER), relation extraction (RE) and event extraction (EE) — in Chinese and English, and ships a fully fine-tuned Chinese-Alpaca-2-13B model trained on roughly 0.4B tokens drawn from close to 50 datasets, part of which was open-sourced through the IEPile corpus. The engineering bet behind it is schema-guided instruction tuning. OneKE builds a hard-negative schema dictionary that maps each label to semantically similar labels, so near-duplicate schemas appear more often in training while overall sample volume drops. A batchified instruction generator then caps how many schemas a single instruction asks for at N, between 4 and 6, spreading the evaluated schema count across instructions and softening the accuracy drop when inference-time schema counts differ from training. Prompts follow a JSON-like dictionary with three fields: 'instruction' (the task and model role), 'schema' (the mutable label list to extract) and 'input' (the source text). The docs recommend embedding schema descriptions and worked examples directly in the instruction — on narrow domains that is the difference between usable and useless output. Training data was normalized and cleaned before use: overlapping text instances with conflicting labels are dropped, and heuristic rules filter non-letter-heavy text, ultra-short unlabeled strings and stopword-dominated samples. Deployment runs through DeepKE-LLM or OpenSPG. With DeepKE-LLM you configure the environment, pull weights, convert data and optionally quantize OneKE (4-bit, for instance) to run on low-power hardware; with OpenSPG you define the schema first and then run extraction jobs. Weights are downloadable from HuggingFace, ModelScope and WiseModel. It suits researchers and NLP engineers building