Stage-0 multiscript pretrained weights for regnetx-det — RegNetX-8GF + D-FINE HybridEncoder + PolylineTransformer, predicting text-line baselines as polylines (B-spline control points), not boxes.

Hub repo: johnlockejrr/regnetx-8gf-polyline-baseline-stage0
Primary file: best_cbad_f1.safetensors (~208 MB)
Paper (Zenodo preprint): doi:10.5281/zenodo.22059333
Recommended operating confidence: 0.40 (CLI --conf 0.4). 0.50 is a close alternative.

This checkpoint is a generic baseline pretrain for downstream fine-tunes (Hebrew/Samaritan, ICDAR cBAD Stage-1, other PAGE/ALTO corpora). It improves on the D-FINE Stage-0 teacher on the same multiscript Arrow validation split.

Sibling checkpoints:

D-FINE Stage-0 teacher: johnlockejrr/dfine-det-large-baseline-stage0 (cbad_f1_max ≈ 0.893)
Hebrew/Samaritan domain fine-tune (older warm-start): johnlockejrr/regnetx-8gf-polyline-hebrew-samaritan (cbad_f1_max ≈ 0.942 on internal val)

Model summary


Architecture	`RegNetXPolylineModel` — timm RegNetX-8GF → HybridEncoder (C3/C4/C5) → PolylineTransformer
Backbone	`regnetx_080.tv2_in1k` (ImageNet-1K V2)
Neck	D-FINE HybridEncoder (`hidden_dim=256`, strides 8/16/32)
Parameters	~54.1 M
Queries	300
Geometry	(K=8) cubic B-spline control points per line + height
Canvas	(1280\times1280) letterbox
Init	ImageNet backbone + shape-matched load from D-FINE Stage-0 (`enc≈543`, `dec≈187`; `input_proj.*.conv.weight` stay RegNetX-specific)
Task	Document baseline / text-line detection → PAGE or ALTO export

Detection is set prediction of scored polylines. Polygonal line environments (BLLA) are applied only at PAGE/ALTO serialize time.

Intended use

Use for

Starting point for Stage-1 / domain fine-tuning on a target script or corpus
Zero-shot probing of baselines on heterogeneous manuscripts (expect domain gap)
Research comparing RegNetX + HybridEncoder vs D-FINE / ConvNeXt polyline siblings

Not for

Claiming official ICDAR cBAD 2019 test numbers from these Stage-0 val scores
Production OCR transcription (lines only; no text recognition)
Layout region detection (paragraphs, tables, illustrations)
Treating cbad_f1 @ conf 0.1 as the quality number (see operating point below)

Training data (Stage-0)

Same multiscript PAGE-XML / ALTO baseline Arrow as dfine-det Stage-0: HTR-United–style open corpora + private packs, compiled with adaptive simplify → uniform cubic B-spline (K=8).

Split	Pages
Train	46 511
Val (holdout)	1 923 (~5%, seed 42)

Corpus mix matches the D-FINE Stage-0 card (29 corpora, multilingual historical scripts). See the dfine-det Stage-0 model card for the full corpus table.

Important: ICDAR cBAD pages appear inside this pretrain mixture. Stage-0 validation F1 is an internal multiscript holdout, not a fair official cBAD 2019 test score. For apples-to-apples cBAD bake-offs, use a separate Stage-1-cBAD fine-tune and the official test split.

Training recipe

Hyperparameter	Value
Init	`regnetx_080.tv2_in1k` + D-FINE Stage-0 encoder/decoder transfer
Optimizer	AdamW; base LR (1\times10^{-4}), backbone (0.2\times)
Schedule	Linear warmup 2 epochs + cosine
Precision	`bf16-mixed`
Effective batch	32 (micro-batch 32 × accum 1 on MI300X)
Epochs (cap)	80; early-stop on `cbad_f1_max` (patience 12, min 15)
Augment	Mild photometric + ±2° rotation (OOB lines dropped)
Match distance	20 px on the 1280 canvas
Monitor	`cbad_f1_max` over conf ∈ {0.1…0.6} with `fast_conf_sweep`
Seed	42

Config reference: configs/baseline_pretrain_stage0_droplet.yaml in the regnetx-det repo.

Hardware: AMD Instinct MI300X on AMD Developer Cloud (ROCm 7.14, PyTorch HIP).

Evaluation (Stage-0 val)

Metric: cBAD-style line F1 — Hungarian matching of densified polylines with mean bidirectional Chamfer cost; TP if cost ≤ 20 px.

Released checkpoint

Metric	Value	Notes
`cbad_f1_max`	≈ 0.929	Best monitored snapshot (`best_0.9293.safetensors`)
Best epoch	32	From bundled TensorBoard logs
Operating conf	≈ 0.40	Sweep winner at best epoch
Precision @ 0.4	≈ 0.930	Same epoch
Recall @ 0.4	≈ 0.929	Same epoch
Mean Chamfer @ best	≈ 4.0 px	Same epoch
Weights file	`best_cbad_f1.safetensors`	Same tensor data as `best_0.9293.safetensors`

Compared to the D-FINE Stage-0 teacher on the same val split: +0.036 absolute cbad_f1_max (0.893 → 0.929).

Fixed conf 0.1 F1 (cbad_f1 on the progress bar) is substantially lower than cbad_f1_max — always sweep (or use ~0.4 for this checkpoint) before comparing systems.

How to use

Install

bash

git clone <regnetx-det-repo>
cd regnetx-det
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Use this project's .venv (not a sibling detector repo).

Download weights

bash

huggingface-cli download johnlockejrr/regnetx-8gf-polyline-baseline-stage0 \
  best_cbad_f1.safetensors \
  --local-dir PRETRAIN/regnetx_stage0

Inference (PAGE XML) — preferred path

bash

regnetx-det infer-page \
  -m PRETRAIN/regnetx_stage0/best_cbad_f1.safetensors \
  -d cuda:0 \
  --format page \
  --conf 0.4 \
  --text-direction horizontal-rl \
  page.jpg

Single-image export with extra NMS/polygon knobs:

bash

regnetx-det infer \
  -w PRETRAIN/regnetx_stage0/best_cbad_f1.safetensors \
  --image page.jpg \
  -o page.xml \
  --format page \
  --conf 0.4 \
  --reading-order rtl

Fine-tune (Stage-1 / domain)

yaml

train:
  load: PRETRAIN/regnetx_stage0/best_cbad_f1.safetensors
  backbone: regnetx_080.tv2_in1k
  pretrained_backbone: false
  lr: 3.0e-5
  backbone_lr_mult: 0.2
  monitor: cbad_f1_max

regnetx-det --config configs/hebrew_samaritan_regnetx_stage0_finetune_balanced.yaml -d cuda:0 train

Hebrew/Samaritan and cBAD Stage-1 YAML templates ship in the regnetx-det repo.

Files in this release

File	Upload to Hub?	Description
`best_cbad_f1.safetensors`	Yes	Recommended Stage-0 weights
`README.md`	Yes	This model card

Limitations

Val F1 is not official cBAD 2019 test performance.
Pretrain includes cBAD pages → do not treat Stage-0 zero-shot cBAD scores as an external bake-off.
(K=8) control points underfit strongly curved / torn lines.
EMA is used during training; released safetensors are the monitored best snapshot.
Some listed corpora may have missing sidecars at compile time; Arrow compile skips unresolved pages.

Citation & credits

Paper (prefer this)

Preprint: regnetx-det: Set-Prediction Polyline Baselines for Historical Documents with HybridEncoder Transfer (Zenodo, Aug 2026).

bibtex

@misc{bors2026regnetxdet,
  author       = {Bors, Teodor},
  title        = {regnetx-det: Set-Prediction Polyline Baselines for Historical Documents with HybridEncoder Transfer},
  year         = {2026},
  month        = aug,
  publisher    = {Zenodo},
  version      = {v1},
  doi          = {10.5281/zenodo.22059333},
  url          = {https://doi.org/10.5281/zenodo.22059333}
}

D-FINE (HybridEncoder / decoder init)

bibtex

@inproceedings{peng2025dfine,
  title     = {D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement},
  author    = {Peng, Yansong and Li, Hebei and Wu, Peixi and Zhang, Yueyi and Sun, Xiaoyan and Wu, Feng},
  booktitle = {The Thirteenth International Conference on Learning Representations},
  year      = {2025},
  url       = {https://arxiv.org/abs/2410.13842}
}

Code: github.com/Peterande/D-FINE
Stage-0 teacher: johnlockejrr/dfine-det-large-baseline-stage0

This Stage-0 checkpoint (weights)

bibtex

@software{regnetx_det_stage0_8gf,
  title   = {regnetx-det Stage-0: RegNetX-8GF Multiscript Polyline Baseline Pretrain},
  author  = {Bors, Teodor},
  year    = {2026},
  note    = {RegNetX-8GF + HybridEncoder + PolylineTransformer; see Zenodo preprint doi:10.5281/zenodo.22059333},
  url     = {https://huggingface.co/johnlockejrr/regnetx-8gf-polyline-baseline-stage0}
}

Acknowledgements

Compute for Stage-0 pretraining was provided via complimentary credits on the AMD Developer Cloud (AMD Instinct™ MI300X) through the AMD AI Developer Program.

Additional notices

Vendored D-FINE modules: see THIRD_PARTY_NOTICES.md in the repo (Apache-2.0).
PAGE polygonalization helpers derive from Kraken/BLLA (Apache-2.0) — used at export, not in the pretrain loss.
Training corpora: respect each dataset's original license. Private packs are not redistributed with this card; only model weights trained with them are released where license allows.

License

Apache License 2.0 for the regnetx-det code and these weights. Downstream users must comply with licenses of any datasets used in further fine-tuning.