Camellia86/Canopy-Jev-27B

Canopy-Jev 27B

Shared context. Parallel decisions. Direct probabilities.

GitHub code & bilingual guide · Full model card

JevBench 87.45% · Hard 75.68% — exceeding Open-Jev-27B-v1.1 by 2.16 / 3.60 percentage points. MMLU-Pro 64.50%. Canopy shares an application-state prefix across isolated question branches and returns probabilities with zero generated answer tokens.

This repository supplies a 25.24 MB adapter and numerical prior for the frozen Qwen3.8-27B text backbone. The custom Canopy-Jev loader supports finite candidate decisions: Choice, binary Noul adaptation and ordered Score adaptation. Evaluation and protocols.

Use

Install the source and tested CUDA environment from GitHub. Runtime: Linux, NVIDIA CUDA, at least 64 GiB free GPU memory. The base snapshot downloads separately.

from pathlib import Path
from huggingface_hub import snapshot_download
from canopy_jev import CanopyJev

snapshot = snapshot_download(
    "Camellia86/Canopy-Jev-27B",
    allow_patterns=["artifacts/*"],
)
model = CanopyJev.from_pretrained(
    Path(snapshot) / "artifacts",
    base_model="Qwen/Qwen3.8-27B",
    revision="1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
)
result = model.predict(
    state="A customer received a damaged order and requests a refund.",
    questions=[{
        "id": "route",
        "type": "choice",
        "question": "Which team should receive the request?",
        "criteria": {"billing": "Refunds and charges", "engineering": "Software defects"},
    }],
)
print(result["questions"])

The artifact manifest verifies the two safe-tensor files and fixed base/tokenizer identity. Source checkpoint: 506410609a6ba6eb422182fddda095b81fa4923011ceaa3717e85a74817237f4. Training and validation records are available in the GitHub model card.

中文

Canopy-Jev（冠层 Jev）：JevBench **87.45%**、Hard **75.68%**，分别超过Open-Jev-27B-v1.1 2.16 / 3.60个百分点；完整MMLU-Pro **64.50%**。

本仓库提供25.24MB适配器与概率先验。一份共享状态支撑多个隔离问题分支，直接输出概率，不生成答案token。源码、安装、时延和完整评测见 GitHub中英文说明。按上方示例下载工件，再通过Canopy-Jev自定义加载器载入固定Qwen基座即可。

License

Weights: Apache-2.0, full license and NOTICE. Original code and documentation: MIT, full terms. Artifact terms · Third-party attribution.

Citation / 引用

Author / 作者：Ruoxi Qiu。

@software{qiu2026canopyjev,
  author  = {Qiu, Ruoxi},
  title   = {Canopy-Jev: Shared-State Numerical Decision Inference},
  year    = {2026},
  version = {0.1.0},
  url     = {https://github.com/Camellia86/Canopy-Jev}
}

CITATION.cff