SLANeXt_wired_bo_ft

Finetuned SLANeXt_wired for wired bordered administrative tables (Bulletin Officiel / legal gazette style). Trained exclusively on public table datasets — no BO ground truth in the training mix.

Base model	PaddlePaddle/SLANeXt_wired
Algorithm	SLANeXt + Vary_VIT_B backbone
Task	Wired table structure recognition → HTML
Resolution	512×512
Training run	bo_ft_gpu_v7
Pipeline	PaddleOCR Table Recognition v2 (wired-only)

Model description

This checkpoint is the table structure component of the TableDetectionRec pipeline. It predicts HTML table structure (<table>, <tr>, <td>, spans) from cropped table images. Companion models (not included here) handle cell detection (RT-DETR-L_wired_table_cell_det) and OCR (PP-OCRv4).

Design choices:

• Wired-only — skips wired/wireless classification (force_wired: true)
• Dynamic orientation — auto-detect 0°/90°/180°/270° at inference
• Encoder freeze — first 10 of 12 Vary_VIT_B blocks frozen during finetune
• Early stopping — patience=5 on validation structure accuracy

Training curriculum

Stage	Name	Dataset(s)	Weight	Train cap	Best val acc	Best epoch	Early stop
1	structure	FinTabNet	40%	20,000	69.3%	15	yes
2	geometry	PubTables (filtered)	30%	40,000	95.0%	6	yes
3	relations	SciTSR	15%	8,000	95.0%	6	yes
4	noise	ICDAR + Marmot	15%	7,000	95.0%	6	yes

Hyperparameters

Parameter	Value
Batch size	8
Learning rate	1e-4 (AdamW + cosine)
Max epochs / stage	20
Val split	5%
Seed	42
Backbone freeze	10 / 12 blocks

Augmentations

Border removal, spacing jitter, ±2° rotation, text stretch, line dropout, brightness/contrast jitter.

Dataset statistics (profiling sample)

Dataset	Samples profiled	Mean cells	Median cells	Max cells	Span fraction
FinTabNet	500	16.0	16	16	0.0
PubTables	500	5.0	5	5	1.0
SciTSR	50	4.0	4	4	0.0
ICDAR	100	19.3	18	30	0.0
Marmot	2,000	16.0	16	16	0.0

Repository layout

inference/
  inference.pdiparams   # Paddle inference weights (~348 MB)
  inference.json        # Model graph
  inference.yml         # PaddleOCR deploy config
  config.json           # Pre/post-processing config
plots/
  01_train_val_loss.svg
  02_table_acc.svg
  04_stage_timeline.svg
metrics.json            # Full training metrics

Usage

TableDetectionRec pipeline

# TableDetectionRec/config/table_recognition.yaml
pipeline:
  force_wired: true
  table_structure_model_dir: finetune/models/best/inference

cd TableDetectionRec
python scripts/02_run_table_recognition.py --run-dir <run_dir> --device gpu:0

PaddleOCR standalone

from paddleocr import TableStructureRecognition

model = TableStructureRecognition(
    model_name="SLANeXt_wired",
    model_dir="./inference",  # path to downloaded inference/ folder
)
output = model.predict(input="table_crop.png", batch_size=1)
for res in output:
    res.print()
    res.save_to_json("./output/res.json")

Download from Hub

huggingface-cli download AvoCahDoe/SLANeXt_wired_bo_ft --local-dir ./SLANeXt_wired_bo_ft

Evaluation metrics (full)

See metrics.json for stage results and dataset profiling stats.

{
  "stages": [
    {
      "stage": 1,
      "name": "structure",
      "best_acc": 0.692999999307,
      "best_epoch": 15,
      "stopped_early": true,
      "n_train": 19000
    },
    {
      "stage": 2,
      "name": "geometry",
      "best_acc": 0.95,
      "best_epoch": 6,
      "stopped_early": true,
      "n_train": 38000
    },
    {
      "stage": 3,
      "name": "relations",
      "best_acc": 0.95,
      "best_epoch": 6,
      "stopped_early": true,
      "n_train": 845
    },
    {
      "stage": 4,
      "name": "noise",
      "best_acc": 0.95,
      "best_epoch": 6,
      "stopped_early": true,
      "n_train": 447
    }
  ],
  "dataset_stats": {
    "fintabnet": {
      "dataset": "fintabnet",
      "n_samples": 500,
      "cell_count_mean": 16.0,
      "cell_count_median": 16.0,
      "cell_count_max": 16,
      "span_fraction": 0.0
    },
    "pubtables": {
      "dataset": "pubtables",
      "n_samples": 500,
      "cell_count_mean": 5.0,
      "cell_count_median": 5.0,
      "cell_count_max": 5,
      "span_fraction": 1.0
    },
    "scitsr": {
      "dataset": "scitsr",
      "n_samples": 50,
      "cell_count_mean": 4.0,
      "cell_count_median": 4.0,
      "cell_count_max": 4,
      "span_fraction": 0.0
    },
    "icdar": {
      "dataset": "icdar",
      "n_samples": 100,
      "cell_count_mean": 19.26,
      "cell_count_median": 18.0,
      "cell_count_max": 30,
      "span_fraction": 0.0
    },
    "marmot": {
      "dataset": "marmot",
      "n_samples": 2000,
      "cell_count_mean": 16.0,
      "cell_count_median": 16.0,
      "cell_count_max": 16,
      "span_fraction": 0.0
    }
  },
  "exported": true,
  "inference_dir": "/workspace/JuridicLayout/TableDetectionRec/finetune/models/best/inference"
}

Limitations

• Trained on public table datasets only; BO-specific layout may still need domain validation.
• Structure model is one piece of the full table pipeline (cells + OCR required for end-to-end extraction).
• Wireless / borderless tables are out of scope (wired-only configuration).

Citation

@misc{slanext_wired_bo_ft,
  author = {AvoCahDoe},
  title = {SLANeXt_wired_bo_ft: Finetuned wired table structure model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AvoCahDoe/SLANeXt_wired_bo_ft}}
}

License

Apache 2.0 (inherits from PaddlePaddle/SLANeXt_wired).