Jev-Style-2B-Decision-v4.2

A diagnostic continuation candidate of 2B v4: 2,213,241,664 parameters including the frozen vision tower, one set of weights, no runtime ensemble. Given a state and typed questions, it returns a probability for every supplied option. It does not generate free-form answers or perform actions.

Public evaluation

The unchanged step-481 HF export scored 46.74 on the complete Decision Index public 0.3 suite: 140,178 scored requests, all successful, no unsupported requests or errors. The archived results were independently rescored on the CPU and reproduced 46.74. The parent v4 public score was 44.37 on the same protocol.

This is a self-run public score, not an official Full/private score or a new vision-board score. The original run used BF16, fast text and CUDA graphs on one RTX 4090, with three disjoint group-hash processes over identical frozen weights. Its concurrent request timings are not serial latency measurements. The exact settings, corpus hashes and original runner hash are preserved in validation/public_run_environment.json. The original tests are not distributed in this repository.

Diagnostic-release disclosure

The pre-registered relative-calibration guard did not pass: candidate ECE 0.0300812254 versus parent ECE 0.0195163271, with an allowed increase of 0.01. The original decision was keep_v4; a separate diagnostic public run was then approved. These weights and their global temperature have not been changed after observing the public score. This release preserves that candidate and the failed guard; it does not retroactively claim a successful pre-registered release decision. See validation/diagnostic_protocol.json and release_config.json.

Run locally

Use Python 3.12 and download this repository, then install its pinned requirements:

python -m pip install -r requirements.txt

For the original CUDA fast path, flash-linear-attention==0.6.0 was used; install the CUDA-compatible build when reproducing that environment. Plain PyTorch fallbacks are correct but slower. The runtime records model loading and graph capture separately. The official maintainers make their own serial-latency measurements.

from jev_style_decision import JevStyleDecision

m = JevStyleDecision('.', device='mps', dtype='float32')  # Apple silicon; CPU also supported
answer = m.decide('The parcel arrived on time and undamaged.', 'Was the delivery successful?')
print(answer['probabilities'])
m.close()

On CUDA use device='cuda', dtype='bfloat16'; the default CUDA fast/graph routing follows the included reference engine. The checked-in src/macjev/ inference modules match the evaluated export byte for byte. The public wrapper adds the callable interface and version label. Generic causal chat-generation runtimes do not reproduce this model's decision probabilities: the renderer, block mask and verdict readout are required.

Protocol and capabilities

• Checkpoint: single step 481 of the v4.2 continuation, preserving its frozen vision tower.
• Global temperature: 0.8813575635811984, the same for choice, noul and score.
• Readout: hidden state at each verdict slot dotted with the tied yes/no embedding-row difference, then softmax.
• Attention: block-causal in full-attention layers; recurrent Gated DeltaNet layers retain their original behavior.
• Complete input budget: 32,768 tokens, at most 255 options; inputs are refused rather than truncated.
• Image processing: native resolution capped at 1,638,400 pixels. Image loading/finite outputs passed a fresh local smoke check; no new full vision benchmark is claimed.

The new packaged runtime passed FP32 MPS checks for all three question types, matching the frozen evaluator's probabilities exactly on the smoke cases. Batched versus single-question probability difference was below 1e-6; overlong inputs were refused. All 1,097 format-fixture questions render identically. See validation/ for the protocol, scope and results. These checks are not a substitute for a full independent benchmark rerun.

Text-only variants are available as GGUF F16 / Q8_0 / Q4_K_M and MLX BF16 / 8-bit. The MLX BF16 files use FP32 activations; the 8-bit files use native BF16 activations. Each format repository includes its custom runtime, precision settings and usage instructions.

Training lineage and notices

This continues the public v4 lineage, whose training included soft labels distilled from TypeSafe Jev 1.14 through its public API, generated/teacher-written material and upstream training splits. The continuation used internal format/rule training and held-out calibration data; the global temperature was fitted on 2,166 internal calibration rows, not benchmark test items. No original Decision Index test corpus is packaged.

JevStyle is independently maintained by Chaoliang Yan, without TypeSafe affiliation or endorsement. The model inherits the parent model's training-data and provider-term considerations; Apache-2.0 for these weights does not relicense third-party datasets or teacher outputs. Read the parent model card, LICENSE and NOTICE. Code and packaging were prepared with AI coding assistance under the project's direction; numerical claims above are restricted to the recorded runs.

Complete public run artifacts: https://huggingface.co/datasets/chaoliangUNSW/decision-index-results-v4.2