OpenJev-0.6B

OpenJev-0.6B

OpenJev-0.6B is the public model name; this checkpoint has model revision 0.3.0, selected by development NLL and used in the published evaluation. All updated parameters are included: the Qwen LoRA adapter and the independent scalar decision head. The unchanged Qwen base is downloaded at the exact revision below. This is an experimental typed-decision model, with supervised training and per-primitive probability calibration.

Load the complete trained decision model

Python 3.12 is recommended. Install the OpenJev loader with the Qwen dependencies:

pip install 'openjev[qwen] @ git+https://github.com/IamBusy/[email protected]'

from openjev import OpenJevModel

model = OpenJevModel.from_pretrained("IamBusy/OpenJev-0.6B")
result = model.predict(
    state="The recorded color is red.",
    questions={
        "color": {
            "type": "choice",
            "instructions": "Which candidate agrees with the recorded color?",
            "criteria": {"red": "The color is red.", "blue": "The color is blue."},
        }
    },
)
print(result["answers"])

The loader downloads this snapshot and the pinned base, checks the file manifest, and loads the custom scoring head and calibration. It needs no model-service API key. Run with local_files_only=True after the first download for offline use; use revision="<repository commit>" to pin this model snapshot. Pass temperatures={} to request uncalibrated output.

Use the OpenJev loader. A generic PEFT text-generation loader only loads the adapter and omits the decision head and candidate-branch scoring behavior. This checkpoint is not a chat model and does not use generate() for its predictions. The Hub text-classification tag describes its purpose; it is not a claim of compatibility with a generic Transformers classification pipeline or hosted widget.

The measured runtime is MPS on an Apple M3 Pro with BF16; CPU uses FP32. This software version does not implement CUDA acceleration. Base download size is about 1.2 GB; allow several GB of disk and RAM for dependencies and execution.

The model can also be downloaded and served with the openjev-model command. The historical openjev-branch command and implementation imports remain valid.

Files and release identity

The adapter and head hashes are identical to the original GitHub v0.3.0 release. OpenJev software v0.3.2 standardizes public names and SDK/CLI entry points. The trained model revision remains 0.3.0. See the naming and version rules.

Publishing LoRA weights with their base reference is a supported PEFT checkpoint format. A merged full checkpoint is another possible distribution. This repository uses the adapter format and includes the additional head required by OpenJev.

Model and use

The checkpoint contains a Qwen3-0.6B attention-projection LoRA (1,146,880 trainable parameters), an independent 1,024-parameter scalar head and per-primitive calibration. The base model is separate, pinned to revision c1899de289a04d12100db370d81485cdf75e47ca of Qwen/Qwen3-0.6B. Use OpenJev's OpenJevModel loader; this is not a standalone Transformers AutoModel or a normal text-generation LoRA. The scalar head is required.

The head starts from the pretrained Yes-minus-No embedding direction and is trained independently. Prediction bypasses the vocabulary projection and does not autoregressively decode answers. Prefix caches live within a request or a training state group and never cross optimizer updates. Candidates are scored independently and normalized within a question.

Intended use is local research into typed decisions, candidate scoring and shared prefix computation. Training/evaluation include English and Chinese simulated states, but broad multilingual capability is not established. The model has no validated role in consequential autonomous decisions.

Data, training and evaluation

725 training judgments from 486 state groups; 176 development, 167 calibration, 405 fresh test and 448 prior regression judgments. The 60-question reference set is a fixed subset of fresh test. Sources include BANKING77, TweetEval sentiment, BoolQ, ARC-Easy and original refund/incident rule scenarios. Access and shipping scenario families are held out. Rules compute scenario labels; missing evidence is an explicit candidate, not a fabricated probability target.

DeepSeek rendered 48 selected worlds and independently reconstructed their facts. 27 renderings passed verification; rejected cases use templates. The same teacher family verified its own renderings, so this is not independent human review. Canonical facts and accepted narratives are in the public scenario snapshot. Public pretraining contamination and template-specific learning remain possible.

Two epochs, seed 43; 244 updates. Selected solely by development NLL (1.0094 to 0.41665). Training took approximately 636 seconds on an M3 Pro with 36 GB RAM. The configuration and hashes are recorded in the code repository's configs/branch-v03.json and reports/v03/TRAINING_SELECTION.json.

On the fixed 60-question reference subset: initial branch model 40/60, trained branch model 45/60, v0.2 Qwen 39/60, DeepSeek 59/60. On 16 matched candidate-reversal checks, v0.3 was consistent in 16/16 cases; an additional 32 fresh cases also passed. These are bounded checks, not universal correctness claims.

Limitations and regressions

Some banking, access-policy and shipping tasks regress relative to v0.2. Access clearance-gap accuracy is 6.67% on its small fresh subset; shipping-action accuracy is 20%. Scores can be confidently wrong. Calibration applies to the measured mixture only. Independent candidates cannot directly interpret other candidates, which limits set-relative criteria such as “none of the above.”

Architecture and training data changed together, so accuracy gains are not an architecture-only ablation. This is one training seed. A 255-candidate smoke test checks execution and output shape, not general large-label performance.

Warm synthetic timing on the M3 Pro: long state/8 questions/4 candidates about 0.70 s versus v0.2's 1.38 s; short state/1 question about 104 ms versus 55 ms. See RESULTS.md in the model bundle or reports/v03/RESULTS.md in the source. No speed or accuracy parity with Jev is claimed. This is supervised training, not a reproduction of TypeSafe's unpublished RLCD algorithm.

Distribution and licenses

The adapter/head bundle and code are distributed under Apache-2.0. The Qwen base is Apache-2.0 and downloaded separately. Dataset texts retain their original licenses and attribution, including BoolQ CC BY-SA 3.0 and ARC CC BY-SA 4.0; see THIRD_PARTY.md. The model bundle includes no raw benchmark corpus, private provider traces or credentials. No legal conclusion about downstream use is implied by model behavior.

The trained adapter and scalar head are available on Hugging Face and in the original GitHub Release ZIP. Both use identical trained tensor bytes. See Hub loading instructions. The original ZIP SHA-256 is pinned in src/openjev/release_assets.json; each bundle has its own file manifest.