Question 1

Does Jeff replace my large model, or does it still call it?

Accepted Answer

It fronts it. Jeff answers first from a single forward pass, and only queries below its per-adapter confidence threshold go on to the stronger model, with both times counted when that happens. On the grounding task, where the large model is already strong, Jeff passes on 1.7% of queries and finishes level with it. That means two models in your serving path, not one.

Question 2

What does it cost me in memory and latency versus just running the big model?

Accepted Answer

By the README's table the large model sits at 28.6 GB and Jeff adds 1.96 GB with all nine adapters loaded, roughly 6.9% more, while mean decision time falls from 8.1 s to 0.25 s. Memory was measured on a workstation GPU switching adapter per request at a 30 ms median, with the accuracy comparison run on a laptop. Both are single-machine results.

Question 3

Can I train an adapter for my own labels, and does it survive a base model upgrade?

Accepted Answer

The data survives, the weights do not. Adapters are about 41 MB LoRA add-ons for the current base, trained in one epoch on one GPU in half an hour to four hours, and an adapter kit ships the checks the maintainers ran on all nine. The README is explicit that adapters do not carry over between base versions while datasets do, so a base bump means retraining everything you use.

Question 4

What happens to a request that matches none of my adapters?

Accepted Answer

It reaches the untouched base model, which is the zero-shot case. You describe the situation and options in plain words and get calibrated probabilities with no generated text and nothing to parse, but accuracy without an adapter is far lower: the untrained small-model column ranges from 12.5% to 59.4% across the nine tasks. Route anything you care about to an adapter.

Question 5

How do I run Jeff, and what question types does it take?

Accepted Answer

Weights are on Hugging Face under mstrasser, with GGUF builds for llama.cpp, and the numbers behind every claim are on jeffhub.ai with the comparison rebuildable from the reference app. Question types are choice with up to 254 options, noul for yes or no, and score on a scale, served over HTTP or through the TypeScript client. v1.3 is adapter-first with 15 adapters, while the v1.2 models stay available under their old names.

Question 6

What are the license terms on the adapters?

Accepted Answer

They differ per adapter and the README says so explicitly. sanctions and soc are CC BY-NC 4.0 because of their source data, which rules out commercial use for those two, and aml and trading-desk carry data-source terms of their own. Every adapter's page on jeffhub.ai gives its licence, its data, where it goes wrong and how sure it is when it is right, and the adapter kit is there to build your own.

Question 7

What did Jeff-Code measure in a real coding agent?

Accepted Answer

A coding agent built on Pi with two small Jeff adapters, jeff-adapter-code and jeff-adapter-code-router, where Jeff takes routine information-gathering steps itself and decides whether Qwen 3.8-27B needs to think hard. Over 1,242 paired tasks from six benchmarks the pass rate was the same, 62.4 against 62.8 percent, with a paired difference of -0.2 points (95 percent interval -2.6 to +2.1), while each task took 0.68x the baseline's time.

Jeff — firelex

About this project

Highlights

Questions about this entry

Other listings of this project

More Jev resources like this

From the Jev guides

From the Jev community