Launching today
ReDefiNER by Jungroo AI Labs
Entity extraction using long, complex rules, on a CPU
1 follower
Entity extraction using long, complex rules, on a CPU
1 follower
ReDefiNER is an open-weight 0.8B model for entity extraction controlled by a plain-English prompt. List the entity types, describe each and add rules about what counts, up to about 10,000 tokens; it returns exact spans with a type and a confidence. The prompt is encoded once and cached, and editing it changes what it finds with no retraining. Built on Qwen3.5-0.8B, it is one 0.77 GB int8 file that runs offline on a CPU. Apache 2.0.
Free
Launch Team / Built With
AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
Hi Product Hunt, I’m Cibe, co-founder of Jungroo AI Labs. ReDefiNER is a small open model that extracts details from text by following long, complex rules you write in plain English. It runs on a CPU, so your text never leaves your machine.
The problem: every team means something different by “the important details”. A privacy team wants names, emails and account numbers. A data team wants the customer, product and date range in a user’s question. Fixed-label extractors fit neither, and sending private text to a large AI model means paying per request and moving the data outside.
How it works: you write a prompt with your entity types, a line describing each, and rules on what to leave out, where a value starts and ends, and which type wins when two fit. Prompts of up to about 10,000 tokens were tested. ReDefiNER returns exact spans from the text, each with a type and a confidence. To change what it finds, you edit the prompt and nothing is retrained.
Where it fits:
• Redaction: a first pass that finds names, phone numbers and account numbers before text goes to an LLM or a vendor
• NL2SQL: the first step, pulling the customer, product, region and dates out of a question before the query is built
• Support tickets: order IDs, product names and error codes, so each ticket reaches the right team
• Teams in banks, hospitals and government that can’t send text to an outside API
Under the hood: it is built on Qwen3.5-0.8B with two small heads. One scores every span of up to 16 tokens, and the other matches spans to the types in your prompt. It never generates text, so every output is a piece of your input. The prompt is encoded once and cached, so each sentence after that takes 0.4 to 0.7 s on a 4-core CPU, or 0.05 s on a GPU. The int8 weights are one 0.77 GB file.
How it compares: given the same prompt on the CrossNER AI test, it scores 68.1, which is 4 to 11 points below top AI models like Claude and Gemini (72.4 to 79.4). With the prompt cached, it costs about 9 times less per request than Claude Haiku 5.5 on a 2-vCPU server.
It’s free under Apache 2.0: pip install redefiner, weights on Hugging Face, code on GitHub, with a Python API and a local HTTP server. Thanks to the Qwen team for releasing Qwen3.5-0.8B under Apache 2.0.
I’d like to hear where it breaks on your own text.
Model on Hugging Face: https://huggingface.co/jungroo-ai/redefiner-0.8b
Demo: https://huggingface.co/spaces/jungroo-ai/redefiner-demo