Question 1

Do I need to download model weights before I can try it?

Accepted Answer

Not to look around. Core imports and the --explain path need no weights and no running service, and extras are split so you can install only the HTTP backend, the local inference stack, or the dev requirements. Weights download on first inference if they are not cached.

Question 2

Why do the tuned models cap at 1024 tokens when the base ones take 32k?

Accepted Answer

The model table shows the tuned repositories at 1024 max tokens against 32768 for the upstream KaLM and Qwen3 configurations, and the fine-tuning recipe states 1024-token truncation. Expect truncation if you point a tuned checkpoint at long inputs.

Question 3

Is it actually better than running KaLM or Qwen3 embeddings directly?

Accepted Answer

The README documents input mapping, scoring rules and worked cases against saved Jev reference outputs, and ships configs for both base and tuned checkpoints, but it does not publish accuracy for the tuned models. Benchmark them on your own data before committing.

Question 4

What does fine-tuning an embedding model into a decision model involve?

Accepted Answer

LoRA on q_proj, k_proj and v_proj at rank 64, alpha 32 and dropout 0.05, in BF16 with an effective batch of 512, learning rate 2e-4 and 10 percent warmup. The released models each trained one epoch over the full training split, and a 1.67 million question dataset is published.

Question 5

Which checkpoints are released, and where do I get the data?

Accepted Answer

JevEmbed-KaLM-Embedding-V2.5 and JevEmbed-Qwen3-Embedding-0.6B and 4B on Hugging Face, each with merged weights and a LoRA adapter, plus CLM-v0.1-8B support with its own setup guide. The training data is the JevEmbed-Data dataset with 1.67 million labeled Choice, Score, and Noul questions. All are Apache 2.0.

Question 6

How do I serve a model and run a decision?

Accepted Answer

python -m jevembed with a config and --serve keeps the model loaded and answers POST /v1/systemone, the same endpoint shape as Jev. A config points at a Hugging Face repository ID, and weights download on first inference if they are not cached locally. You can also drive it from Python via JevEmbed(config=...).evaluate(request), and an external /v1/embeddings service works as a backend.

Question 7

What is in the benchmark and playground modules?

Accepted Answer

The benchmark module uses versioned multi-file JSONL cases with unified Choice, Score, and Noul evaluation and structured JSON results, emitting per-model results, a predictions file, and a comparison summary. The playground runs at http://127.0.0.1:8000/playground/ after python -m jevembed --playground, with model comparison, JSON inspection, and a Game Lab of model-controlled pixel games.

JevEmbed

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community