AutomatosX/AXDocLayout-M-v1.5-CUDA-TF32

AX-DocLayout-M v1.5 - H5.1 CUDA / cuDNN TF32

This is the owner-selected H51-title final-1024 EMA model. It continues H5 finance fine-tuning with a paired Title-sampling study. It is a separate Heron-derived model identity, not D5/D6/H7, not the public eleven-class v1.0 model, and not an ax-doclayout Python package release. The original v1.4 weights and public v1.0 loader pin remain unchanged.

Decoded RGB page raster in; layout boxes out. Native classes0-16 map to AX PDF SDK IDs1-17 via sdk-taxonomy.json (ax-document-unified-25/1). SDK18-25 are unsupported. This model does not open PDFs, perform OCR or infer reading order. Accuracy retention of extra SDK classes12-17 is unmeasured.

Exact artifact and qualified runtime

• FP32 safetensors SHA256: d0949704220ee50d94359cb6417e3481ac4ef5078a9f63795989e9f5d76e9b7e.
• H51-title final1024 checkpoint SHA256: 44b75b2ca14d7d4756ca51d402d470d069f17ee99145717e1994549b32ef87d2.
• Upstream: docling-project/docling-layout-heron at immutable revision 8f39ad3c0b4c58e9c2d2c84a38465abf757272d8, Apache-2.0.
• Native 640x640 direct resize, original hash-pinned image processor. Do not substitute D5's1024letterbox or H7's1696letterbox preprocessing.
• Tested CUDA: PyTorch2.12.1+cu129, CUDA12.9, Transformers4.53.3, RTX PRO6000Blackwell. FP32 storage/activations, no autocast or quantization.
• cuDNN TF32 allowed; CUDA matmul TF32 disabled. The suffix means cuDNN permission, not all-matmul TF32. No ONNX or TensorRT artifact is included.
• The exact selected checkpoint has existing three-panel CUDA evidence and exact exported-tensor reload evidence. The packaged example is CPU-qualified; it was not freshly rerun on CUDA for this publication. No GPU was restarted.

Measured results and accepted tradeoffs

All AP values are percentages on previously exposed development panels.

Panel	H5	Matched control	Selected H5.1 / v1.5
General AP (6489pages)	70.6572	71.1214	71.5327
Financial AP (1745pages)	61.5113	61.5445	61.7331
FinTab AP (256pages)	91.5856	91.7371	91.5774
General Title AP	35.9637	37.0977	38.7453
General Title precision at0.30/IoU0.50	38.9333	39.5833	36.2069
General Title recall at0.30/IoU0.50	48.8294	50.8361	56.1873
FinTab AP90	89.0473	89.0155	88.4481

Title AP treatment-control improves1.6476pp, with a paired family-bootstrap95% interval[0.4851,2.8212]. Compared with H5 at the fixed threshold,22additional true Titles accompany67additional false Titles. The frozen numerical screen failed, including the2pp Title AP target, Title precision retention and table AP90 retention. The owner accepted these tradeoffs and explicitly selected publication; the original screen has not been reclassified as passed.

Financial is a subset of General, not independent HKEX evidence. Family independence, independent final-test quality, second-seed reproducibility and extra-class accuracy are not established. This is not a claim of overall superiority to D5, D6 or H7; their differently defined panels cannot be ranked by subtracting historical scores. See evaluation.json and the frozen screen.

Installation and inference

Use Python3.12/3.13. Install CUDA PyTorch separately for the recorded runtime:

python -m pip install torch==2.12.1 --index-url https://download.pytorch.org/whl/cu129
python -m pip install -r requirements.txt

Download this private repository with an authorized Hugging Face account. Pin the immutable commit from the release receipt, or resolve tag v1.5 to that commit and use hf download --revision <COMMIT>. The example fails closed on core hash or runtime mismatches and never falls back silently from CUDA.

python predict_image.py --model-dir . --image page.png --output layout.json --device cuda

Use --device cpu explicitly for CPU. Boxes use original raster pixel xyxy coordinates and SDK class IDs1-17. The default score threshold is0.30; it is an operating point, not COCO AP/AR100. Keep all17classes when interpreting output. Do not load this artifact through the eleven-class v1.0 receipt-pinned loader.

The bundle contains inference files, aggregate numeric evidence, attribution and hash receipts only. No training documents, pixels, per-page predictions, optimizer states, credentials or private training implementation are included.