This model is a conversion of MoritzLaurer/roberta-base-zeroshot-v2.0-c to ONNX format using the Optimum library.
OpenJev-0.6B is an open-weight model for zero-shot classification from Zhiwei Li, released under MIT License. Its published files total 1.2 GB.
An open 0.6B "System One" decision model aligned with the public Jev HTTP contract: state + questions in, complete calibrated probability distributions out, zero autoregressive decoding.
Parameters—
Context—
Weights1.2 GB
Licensemit
AccessOpen weights
Monthly Downloads—
Model Card
By Zhiwei Li, published under mit, revision f7501d08f822.
An open 0.6B "System One" decision model aligned with the public Jev HTTP contract: state + questions in, complete calibrated probability distributions out, zero autoregressive decoding. GLM-5.3-flash structured teacher distillation, CE + Brier, best-step selection and calibration-split temperature fitting. One RTX 4080 SUPER, ~65 min, full recipe in the repo. - Results (fixed held-out websites): choice top-1-in-positives 0.380 (test) / 0.500 (ood) vs uniform 0.089; noul accuracy 0.84 / Brier 0.142; real Jev-1.13 reaches 0.543 / 0.610 on the same questions. Or serve the Jev-compatible HTTP API: python -m openjev.serve --checkpoint-dir --port 8123. Full pipeline (data synthesis → streaming…
Read Zhiwei Li's full model card
An open 0.6B "System One" decision model aligned with the public Jev HTTP contract:
state + questions in, complete calibrated probability distributions out, zero autoregressive decoding.
- Backbone: Qwen3-0.6B + a ~2M-parameter shared scalar decision head (one logit per candidate path).
- Training: Mind2Web (axtree-cleaned) web element-selection records, website-disjoint splits, GLM-5.3-flash structured teacher distillation, CE + Brier, best-step selection and calibration-split temperature fitting. One RTX 4080 SUPER, ~65 min, full recipe in the repo.
- Results (fixed held-out websites): choice top-1-in-positives 0.380 (test) / 0.500 (ood) vs uniform 0.089; noul accuracy 0.84 / Brier 0.142; real Jev-1.13 reaches 0.543 / 0.610 on the same questions.
Usage
from openjev.predict import Predictor
p = Predictor("AndeyTait/OpenJev-0.6B") # or a local checkout dir
p.decide("Task: open settings. Elements: [e1] link Settings, [e2] button Save", {
"next": {"type": "choice", "instructions": "Which element next?",
"criteria": {"e1": "link Settings", "e2": "button Save"}},
"is_link": {"type": "noul", "instructions": "e1 is a link that opens settings."},
})
Or serve the Jev-compatible HTTP API: python -m openjev.serve --checkpoint-dir <dir> --port 8123.
Full pipeline (data synthesis → streaming distillation → training → evaluation → serving): see the OpenJev repository. Mind2Web annotations used under their research license; teacher labels generated via OpenRouter.
Configuration
- Hidden size
- 1,024
Identity and Version
- Repository
- AndeyTait/OpenJev-0.6B
- Publisher
- Zhiwei Li
- Task
- Zero-shot classification
- Modality
- Text
- Library
- transformers
- Parameters
- Not stated by the source
- Languages
- en
- Revision
- f7501d08f822d285ad059c3276e4b67eadd77642
- First published
- 2026-09-19
- Last updated
- 2026-09-19
Files and Weights
14 files, 1.2 GB in total. The weights are 1 file totalling 1.2 GB in safetensors.
Weights1 file · 1.2 GB
Configuration4 files · 3.9 KB
Tokenizer4 files · 15.9 MB
Documentation3 files · 27.1 KB
Repository2 files · 3.1 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| best.safetensors | Weights | 1.2 GB | b00c6ed649df |
| backbone/config.json | Configuration | 726 B | — |
| backbone/generation_config.json | Configuration | 239 B | — |
| config.json | Configuration | 755 B | — |
| train_log.json | Configuration | 2.2 KB | — |
| README.md | Documentation | 1.8 KB | — |
| backbone/LICENSE | Documentation | 11.3 KB | — |
| backbone/README.md | Documentation | 14.0 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| backbone/.gitattributes | Repository | 1.6 KB | — |
| backbone/merges.txt | Tokenizer | 1.7 MB | — |
| backbone/tokenizer.json | Tokenizer | 11.4 MB | aeb13307a71a |
| backbone/tokenizer_config.json | Tokenizer | 9.7 KB | — |
| backbone/vocab.json | Tokenizer | 2.8 MB | — |
License and Download
- License
- mit
- Access
- Open weights, no gate
- Download size
- 1.2 GB
Download from Zhiwei Li
Released by Zhiwei Li through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen3-0.6B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 1.2 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About OpenJev-0.6B
Can I use OpenJev-0.6B commercially?
Yes. OpenJev-0.6B is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.
Similar Models
https://huggingface.co/facebook/bart-large-mnli with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).
distilbart-mnli is the distilled version of bart-large-mnli created using the No Teacher Distillation technique proposed for BART summarisation by Huggingface, here. We just copy alternating layers from bart-large-mnli and finetune more on the same data. This is a very simple and effective technique, as we can see the performance drop is very little. Detailed performace trade-offs will be posted in this sheet. If you want to train these models yourself, clone the distillbart-mnli repo and follow the steps below Clone and install transformers from source Download MNLI data Create student model Start fine-tuning You can find the logs of these trained models in this wandb project.
distilbart-mnli is the distilled version of bart-large-mnli created using the No Teacher Distillation technique proposed for BART summarisation by Huggingface, here. We just copy alternating layers from bart-large-mnli and finetune more on the same data. This is a very simple and effective technique, as we can see the performance drop is very little. Detailed performace trade-offs will be posted in this sheet. If you want to train these models yourself, clone the distillbart-mnli repo and follow the steps below Clone and install transformers from source Download MNLI data Create student model Start fine-tuning You can find the logs of these trained models in this wandb project.
https://huggingface.co/typeform/mobilebert-uncased-mnli with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).
https://huggingface.co/cross-encoder/nli-deberta-v3-xsmall with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).