JDI-Decision-V1

中文 | English

Organization: JD.com (京东)

JDI-Decision-V1 is a multimodal decision system from JD.com, based on Qwen3.8-27B with D3 decision-architecture adaptations. It provides structured choices and probabilities for text and images. Its internal modules include a vision encoder, a frozen PaddleOCR v5 text recognition component from Baidu’s PaddlePaddle team, a decision backbone and an FP32 readout.

The internal OCR and vision-encoding branches run concurrently on one GPU. The reported vision scores evaluate the complete JDI-Decision-V1 system, and end-to-end request latency includes the OCR component. Every benchmark uses the same system configuration and inference procedure.

Architecture and inference flow

Standard image input handling decodes images, converts them to RGB, and resizes them to meet the pixel cap before the two internal parallel branches. The vision encoder extracts image features; the frozen PaddleOCR v5 component recognizes text, which is added through state.image_text_transcription alongside the original state. The decision backbone processes visual features together with the text, questions and options, then uses an FP32 readout to produce structured answers and option probabilities.

Each request runs on one GPU. Visual features are reused only within that request; OCR text supplements rather than replaces the image information. Base architecture and decision-runtime details are provided in the reproduction guide.

Run

Use Python 3.12 on a CUDA GPU. From the downloaded repository:

# First follow REPRODUCE.md to install the environment and prepare example.png.
source .venv-ocr/bin/activate
CUDA_VISIBLE_DEVICES=0 python - <<'PY'
from ocr_inference import ComboEngine
engine = ComboEngine()
try:
    response, raw = engine(
        state={},
        questions={"answer": {"type": "choice", "instructions": "What does the sign say?",
                              "criteria": {"A": "STOP", "B": "GO"}}},
        images=["example.png"],
    )
    print(response)
finally:
    engine.close()
PY

The setup script retrieves the exact evaluated OCR model archives from Paddle's official hosting and verifies SHA256 before extraction. Sources and hashes are in ocr_inference/ocr_sources.json. OCR weights and Paddle build headers retain their upstream licensing terms.

Public vision evaluation

Decision Index Vision public suite: 79.7313, 12,210/12,210 successful requests across 11 benchmarks. This is a self-reported public score, not a confirmed leaderboard rank or Full score. Private evaluation and official RTX PRO 6000 latency validation are pending.

The run used two H200 workers, each with one complete system replica and one request at a time. Each request stayed on one GPU. Request latency includes decoding, OCR, vision encoding and the decision forward pass; loading, integrity hashing and profiler runs are excluded.

Results and environment metadata: results dataset. The complete public results include OCR transcriptions, answers, probabilities, payload hashes and timings. results.jsonl is byte-identical to the evaluated run; see export-manifest.json for its SHA256.

See REPRODUCE.md for suite reconstruction and evaluation. ocr_inference.ComboEngine is the reproduction entry point for the complete JDI-Decision-V1 system; the release package passed regression comparisons against the original evaluation implementation. inference.D3Engine is the internal decision-model entry point and does not include the complete system's OCR processing.

AutoModel loading

The complete system also supports AutoModel.from_pretrained(..., trust_remote_code=True), including its frozen internal PaddleOCR v5 component. Install the pinned dependencies first, then follow section 5 of the reproduction guide for Hub-ID or local-directory loading. Six H200 requests matched answers, probabilities and OCR text exactly; weights and public evaluation results are unchanged.