Bring Your Own Schema: How We Built kniv v7
Blog post from Hugging Face
Kniv v7 is an Apache-2.0 family of six local English information-extraction models, ranging from 55M to 538M parameters, designed to combine fixed linguistic analysis with schema-driven extraction of user-defined entities, relations, polarity, modality, speaker attribution, and time without retraining or LLM API calls. Built for document and conversation windows rather than isolated sentences, it uses shared encoders with plain, label, and entity-marker passes, allowing it to resolve coreference and identify facts whose evidence spans multiple sentences. The project found that broad and varied supervision, positive-unlabelled learning for incomplete labels, and calibration of thresholds mattered more than increasingly complex prediction heads; it trained two specialized large teacher models, UltraBERT for open schemas and longer inputs and DeBERTa for parsing and fixed-schema relations, then distilled their combined outputs into smaller variants. Models are distributed with PyTorch and ONNX versions including optimized INT8 graphs, enabling CPU inference with no external data transfer, and reported benchmarks show competitive linguistic and relation-extraction performance, with the authors claiming substantial speed and cost advantages over hosted GPT-6 extraction for a memory-oriented workload. The release nevertheless notes limitations involving cross-window linking, closely related relation labels, conversation attribution, negation and reported speech, as well as weaker coreference and entity typing in smaller models.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| LLM | 7 | 3,355 | 742 | 191 | -49% |
| AI Model Fine-tuning | 2 | 613 | 130 | 56 | -29% |
| Vector Search | 2 | 1,097 | 201 | 87 | -45% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.