JevEmbed
Turn embedding models into Choice, Score, and Noul: JevEmbed-Data, LoRA releases on Hugging Face, benchmark runner, playground, HTTP /v1/systemone.
Overview
JevEmbed (github.com/HITsz-TMG/JevEmbed, Apache-2.0) is HITsz-TMG's toolkit to turn embedding models into System One shaped Choice, Score, and Noul outputs without autoregressive decoding. Hugging Face collection HIT-TMG/lychee-jevembed ships merged and LoRA adapters such as JevEmbed-Qwen3-Embedding-0.6B and 4B; dataset HIT-TMG/JevEmbed-Data lists about 1.67 million labeled questions. CLI python -m jevembed runs inference; --serve exposes POST /v1/systemone; --playground opens an interactive UI. Docs cover synthesis, LoRA fine-tune on 1,601,157 training questions, benchmark JSONL, and HTTP backends including external /v1/embeddings. README reports held-out JevEmbed-Data test accuracy up to 85.86% overall for the 4B fine-tune versus 36.29% base Qwen3-Embedding-4B. Independent of TypeSafe; not claiming official affiliation. Public GitHub counts were 60 stars when this listing was drafted.
Problem: Embedding models return vectors, not calibrated Choice, Score, or Noul decisions, so teams either call hosted System One or bolt chat JSON on top of retrieval.
Built for: ML engineers and researchers who want open-weight embedding backends fine-tuned into Jev-style decisions with synthesis, LoRA training, HTTP serve, playground, and published JevEmbed-Data benchmarks.
First indexed on Jev Directory: 2026-09-30
Creator and team
- Name
- HITsz-TMG
- Handle
- @HITsz-TMG
How Jev is used
- Role in the product flow
- Embedding forward pass plus decision heads for Choice, Score, and Noul; training, benchmark, HTTP API, and playground exploration
- Primitives
- ChoiceScoreNoul
- State in
- System One request JSON with state text and questions map per examples/ and docs/examples-and-scoring.md; configs select base or JevEmbed fine-tuned weights.
- Decision out
- Probability distributions and labels matching System One response shape from local inference or HTTP /v1/systemone.
- Load ModelConfig YAML for a base encoder or HIT-TMG JevEmbed release
- JevEmbed.evaluate or CLI --input runs Choice, Score, or Noul heads on embeddings
- jevembed-benchmark writes per-model results.json and summary.json on JSONL suites
- LoRA train on JevEmbed-Data JSONL with documented BF16 batch 512 and rank 64 recipe
- python -m jevembed --serve listens for curl POST /v1/systemone clients
JevEmbed targets embedding checkpoints, not Apple Neural Engine ports. mizorewww-laya-coreml runs open Laya on Core ML with ANE bundles and Snake demos. togethercomputer-tev1 fine-tunes Qwen for letter Choice on Together. logan-markewich-jeff self-hosts GLiFormer System One HTTP. featherless-simple-jev assembles logits from arbitrary HF models without HIT-TMG's JevEmbed-Data recipe. Use JevEmbed when the story is turn my embedder into decisions with published HF weights and academic citations. Cite HITsz-TMG/JevEmbed README, docs/benchmark.md, docs/training.md, docs/http.md, docs/playground.md, and reports/JEVEMBED_DATA_TEST.md only.
Sourced performance claims
- README JevEmbed-Data test table: JevEmbed-Qwen3-Embedding-4B reaches 90.31% Choice, 73.41% Score, 95.88% Noul, 85.86% overall on 64,110 hard-labeled questions of 66,482 total test rows.Source: github.com/HITsz-TMG/JevEmbed README and reports/JEVEMBED_DATA_TEST.md
- Fine-tuning recipe: 1 epoch on 1,601,157 training questions, BF16, effective batch 512, LoRA rank 64, learning rate 2e-4, 1024 token truncation.Source: github.com/HITsz-TMG/JevEmbed README LoRA fine-tuning
- Dataset HIT-TMG/JevEmbed-Data documents about 1.67 million labeled Choice, Score, and Noul questions for training.Source: github.com/HITsz-TMG/JevEmbed README News and Hugging Face dataset card
- Public GitHub repo HITsz-TMG/JevEmbed had 60 stars when this listing was drafted.Source: GitHub API 2026-09-30
Features and stack
Features
- Choice, Score, and Noul heads on multiple embedding backends
- Hugging Face JevEmbed model collection and JevEmbed-Data dataset
- Supervised synthesis pipeline for new question schemas
- LoRA fine-tune with merged releases and adapter cards
- HTTP POST /v1/systemone server and interactive playground
- jevembed-benchmark JSONL runner with summary comparison
Stack
- Python 3.10 to 3.12
- Sentence Transformers and configurable backends
- Hugging Face weights under HIT-TMG
- Apache-2.0
Pricing: Open source toolkit; inference cost is your GPU, electricity, and optional external embedding HTTP fees documented in docs/http.md.
Links
FAQ
- Is JevEmbed a TypeSafe or System One host?
- It implements System One shaped requests and responses on embedding models. It is an independent HITsz-TMG research toolkit, not a TypeSafe product.
- How do I try a released model quickly?
- README Quick start: python -m jevembed with a configs/jevembed-qwen3-embedding-0.6b.yaml config, or --playground for the browser UI.
- What data was used for the published accuracy jumps?
- README cites JevEmbed-Data with 1,601,157 training questions and the 66,482 question test split in reports/JEVEMBED_DATA_TEST.md.
- How is this different from Laya-CoreML?
- Laya-CoreML ports Convai Laya to Apple Silicon Core ML. JevEmbed fine-tunes general embedding encoders on JevEmbed-Data and serves /v1/systemone from Python.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.