Polish OCR LoRA (BROKEN - DO NOT USE IN PRODUCTION)

WARNING: This model is broken and produces garbage output. It is uploaded for archival/experimental purposes only.

Status: FAILED EXPERIMENT

This is a LoRA adapter fine-tuned on PaddleOCR-VL for Polish OCR tasks. The training did not converge properly and the model outputs are unreliable.

Known Issues

  • Model produces hallucinated text
  • Poor accuracy on Polish characters (especially: a, e, o, s, z, n, c, l)
  • Inconsistent output quality
  • May output repetitive or nonsensical text

Model Details

  • Base Model: PaddlePaddle/PaddleOCR-VL
  • Adapter Type: LoRA (r=16, alpha=32)
  • Target Modules: q_proj, v_proj, o_proj, k_proj
  • Task: Polish document OCR (attempted)
  • Language: Polish
  • Training Data: Synthetic Polish OCR dataset (10k samples)
  • Framework: PEFT 0.18.0

Why Upload a Broken Model?

  1. Transparency: To document what doesn't work
  2. Reproducibility: Others can learn from this failure
  3. Baseline: Can be used as a negative example for benchmarking

Training Configuration

LoRA rank: 16
LoRA alpha: 32
Dropout: 0.05
Checkpoints: 846 steps

Do Not Use For

  • Production OCR systems
  • Any task requiring accurate text extraction
  • Anything where correctness matters

License

Apache 2.0

Framework Versions

  • PEFT 0.18.0
  • Transformers (latest at time of training)
Downloads last month
9
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kacperwikiel/polish-ocr-lora-broken

Adapter
(6)
this model