Jev directory
CommunityJev Models & Reimplementations
FIS-Jev
A LoRA adapter for Qwen3-Omni-30B-A3B-Instruct that scores a therapist's spoken response on the eight Facilitative Interpersonal Skills items, returning a 1-5 distribution and expected value per item from video, audio, and transcript in one forward pass.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- LoRA r=16, alpha=32 on q, k, v and o projections of all 48 decoder layers; 13.4M parameters.
- One forward pass reads the logits of tokens 1-5 at 8 slots, giving a 1-5 distribution per FIS item.
- FIS total correlates with the mean of 3 raters at Pearson r 0.71 on 230 held-out responses, versus 0.54 zero-shot.
- Runs on one 80 GB GPU (about 64 GB in bf16); roughly 130-300 ms per pass on a single H100.
- Trained on 996 FIS responses split by participant; the training data is not included and is not public.
Questions about this entry
Short answers drawn from the linked project or page.
Install the requirements from the fis-jev code repository, then run fis_jev/infer.py with the base model path, adapter.safetensors, the response video, a transcript file, and the client prompt and client-says context. The adapter uses its own LoRA layer rather than PEFT.
One 80 GB GPU, with the model taking about 64 GB in bf16. The model card reports testing with torch 2.14.1 and transformers 5.19.0. A single pass takes roughly 130-300 ms on one H100 after warm-up.
JSON with a score (1-5) and a probs distribution over 1-5 for each of the eight items, plus fis_total, the mean of the eight expected values, along with input_tokens and forward_ms. It does not generate text.
On 230 held-out responses, FIS total correlates with the mean of three raters at Pearson r 0.71, against 0.54 for the zero-shot base model. On 178 responses with three raters, the model's mean agreement with the other two raters is 0.64, versus 0.62 for a single human rater.
The model card reports per-item expected calibration error of 0.03-0.15 with no post-hoc calibration, measured in the same setup trained on the train split only. The 1-5 distribution returned per item serves as the confidence.
Eight items in slot order: verbal fluency, hope, persuasiveness, emotional expression, warmth, empathy, alliance bond, and rupture repair. Per-item correlation with the three-rater mean ranges from 0.55 for warmth to 0.75 for hope.
It is intended for research and for practice feedback in training settings, not for clinical, hiring, or certification decisions. It is trained and tested only on the FIS task, multi-turn sessions are not validated, and accuracy on automatic speech recognition transcripts is untested.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Hugging Face dataset of 1,245,391 synthetic decision-modelling examples carrying Jev teacher probability distributions, used for Claes training. Access is gated and requires manual approval of the repository conditions.
Jev Models & Reimplementations#open-models#calibration#evaluation
Community
A preregistered measurement of a Jev-style interface on frozen Qwen3-4B: read the option letter's logits instead of generating grammar-constrained JSON.
Jev Models & Reimplementations#community#python#local
Community
r-ms/mini-jevPython
r-ms/mini-jevPythonA Hugging Face checkpoint that fine-tunes UltraBERT-380M into a Jev-style decision model: it reads a state, a question and a list of options and picks one in a single forward pass. Reports 72.6% mean accuracy over 13 datasets on the test split.
Jev Models & Reimplementations#open-models#classification#local
Community
A trained example bank for @ruvector/typesafe holding 9,993 labelled Banking77 utterances across 77 banking intents with frozen split assignments. Heads and temperature calibration are refit from the bank on load; no weights are stored.
Jev Models & Reimplementations#open-models#classification#local
Community
MLX precision variants of Jev-Style-2B-Decision-v4.2, a 1.88B-parameter text-only decision model for Apple silicon. Ships bf16 and 8-bit weights with a custom runtime that returns a probability for every supplied option.
Jev Models & Reimplementations#open-models#local#python
Community
Zuse is a family of three decision models (4B, 9B, 27B) that train a low-rank adapter and pointer head on frozen open-weight bases for 27 European languages. The 27B model reaches 87.5% accuracy with pooled expected calibration error of 0.016 on an 11-source European suite.
Jev Models & ReimplementationsArticle#arxiv#open-models#calibration
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.