Labelling text at volume is the oldest use for a classifier, and the one where a decision model is easiest to try. You give it a list of labels, and each text comes back with one label and a confidence value. No training run is needed to start, so you can test a taxonomy in an afternoon. The examples in this hub run on Jev and other decision models, and Laya appears in many, including versions fine-tuned for Arabic and Burmese. Examples include 1kpapers by Hassan, which sorts AI research papers into topics, AkashPriyadarshii/jev-curate, a dataset sifter for Parquet and JSONL rows, and a DuckDB extension that classifies rows of a CSV. Before you copy one, check how many labels it can hold in one question, how it was scored against human labels, and which languages it was tested on. Also check what it does with the low-confidence rows. Fine-tuned entries show what a small amount of local data can change.
20 entries
Showing 20 of 20
Fine-tunes Sentence Transformer embeddings plus a lightweight head for high-accuracy classification from a handful of labeled examples, with no prompting. The standard cheap alternative to an LLM classification call when the label set is fixed. Star count is GitHub's rounded display figure.
Rust and Python dataset sifter that streams Parquet and JSONL rows past Jev typed decisions at 1,500+ records per second, writing passed and rejected rows to disk. Uses speculative fan-out and calibrated post-training rubrics.
Recursive Choice over a taxonomy. Jev can only list 255 options in one choice question, so jev-tree walks a JSON shape and calls it once per level or partition, letting you select among thousands of leaves.
A frozen 64-case benchmark of Feishu-style workplace messages, scoring Jev and Laya on a single choice question and on a four-question workflow. Jev scored 100 and 98.44 percent correct, Laya scored 31.25 and 28.12 percent, on the same requests.
Zero-shot classification benchmark across four datasets: Jev leads decisively on topic tasks, and Laya's advertised edge over Jev does not reproduce, tying on emotion and trailing elsewhere.
A zero-shot benchmark of Jev, self-hosted Kev-9B and self-hosted Laya on 171 Moroccan Darija reviews in Arabic script and Arabizi. Jev scored 79.5 percent accuracy, Kev-9B 71.9 percent, and Laya 53.8 percent, tying an always-positive baseline.
Hassan used Jev to sort 1,018 AI research papers into one of 24 topics for $0.08 in total, with a 256 ms median latency per paper, after DeepSeek V4 Flash wrote the summaries. He was still running evals before swapping them in for the topics on 1kpapers.com.
Hassan (@nutlope)Open1kpapers, AI research papers classified by Jev on 1kpapers.com286k viewsCost $0.08Latency 256 ms median
A DuckDB extension that runs Jev classification on rows of any CSV, Parquet file or DuckDB table, taking about 10 seconds for 1,000 rows.
Hamilton Ulmer (@hamiltonulmer)OpenA DuckDB extension for Jev classification (X) on x.com181k viewsPer 1k rows ~10s
Ran 14 yes/no questions per post, such as whether the hook opens a loop or whether a claim is proven, over 100,000 X posts in 20.4 seconds for 67 cents, then compared which hook types actually beat the viral baseline.
Movez (@0xMovez)OpenA viral post analyzer for 100,000 X posts (X) on x.com82k viewsFull run 20.4s, $0.67vs Opus 5 per post ~680x cheaper
A personal research dashboard where Jev labels saved X posts across 8 dimensions, including topic, hook and writing style, so the poster can filter and compare engagement. 1,315 posts cost about $0.086 in estimated model cost.
Yum⋆₊˚ (@yuhasbeentaken)OpenA research dashboard for X posts labeled by Jev (X) on x.com70k viewsCost ~$0.086Posts labeled 1,315
A video demo of a prototype that feeds a live clinical consultation transcript to Jev, which traverses a medical ontology to classify symptoms, update the differential diagnosis, suggest treatments, and flag red flags in real time.
Chris McMaster (@rheum_ai)57k viewsOpenLive clinical-transcript classification prototype (X) on x.com
Paste up to 10,000 words and see the generated-sounding sentences worth a second look. Free and no login.
Jozef (@jozef_gherman)OpenJev Detector: Free AI Slop Detector on jevdetector.com36k viewsScan speed ~2s/10k words
A video demo combines OCR with Jev to build an image classifier that categorized about 900 images in 40 seconds.
Fayaz Ahmed (@fayazara)OpenOCR-powered image classification (X) on x.com34k viewsThroughput 900 in 40s
A video demo of Jev, now available through the Venice API, classifying 24,000 Hacker News posts into 12 categories in about 2 minutes.
Sabrina (@sabrinaesaquino)14k viewsOpenClassifying 24,000 Hacker News posts on the Venice API (X) on x.com
Laya's multilingual checkpoint fine-tuned on 701 Burmese SIB-200 examples for seven-way topic classification, reaching 0.716 test accuracy versus a 0.441 zero-shot baseline, with temperature-scaled calibrated probabilities.
Laya fine-tuned for Arabic NLU on laya-multilingual, covering intent classification, natural language inference and offensive-language detection. On MASSIVE ar-SA intent classification it reached 81.6 percent accuracy against 38.6 percent for stock Laya.
Runs early-access Jev over 24 Norwegian submissions to the 2022 hearing on resource rent tax for salmon farming, reporting agreement with a frontier model's labels, calibration, cost and latency. Compared against DeepSeek V4.1 Flash with and without reasoning.
Uses the confidence value returned with Choice and Score answers to decide when to act automatically and when to escalate.
Classifies into a deep taxonomy by walking the tree one Choice question at a time instead of flattening every leaf into one list.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.
A typical job streams rows past a few typed questions and writes the answers to a table. AkashPriyadarshii/jev-curate writes passed and rejected rows to separate files. The TypeSafe AI cookbook on classification using confidence shows the key step: act automatically on high-confidence rows and send the rest to a person or a larger model. That split keeps people on the hard cases only.
Choice fits single-label work. Score fits ratings such as formality or sentiment. Noul fits yes or no attributes, and some entries ask many of them per text, like the viral post analyzer by Movez, which runs a set of yes or no questions on each X post. Which primitive you pick changes the downstream table, so decide early whether a column is a label, a rating or a flag.
Large label sets are the limit. One choice question takes at most 255 options, so reachjalil/jev-tree and the TypeSafe AI cookbook on hierarchical classification walk a taxonomy one level at a time. Language matters too. Jev vs Laya on Moroccan Darija sentiment tests zero-shot classification on Arabic script and Arabizi. A taxonomy that works in English needs a fresh test in another language.
Yes. You describe the labels in plain words and ask a choice question. Several entries here test this zero-shot. Check them on your own labeled sample, since results vary by task and language.
A choice question takes up to 255 options. For larger taxonomies, reachjalil/jev-tree and the hierarchical classification cookbook ask one level at a time.
Send them to a person or a larger model. The TypeSafe AI cookbook on classification using confidence shows how to set the cutoff so clear rows are handled automatically.
