CircuitOCR LoRA Weights
LoRA fine-tuning weights for PaddleOCR-VL-0.9B on circuit schematic OCR.
Available Checkpoints
| Version | File | CompF1 | NED ↓ | Description |
|---|---|---|---|---|
| v2 (Phase 1) ★ | lora_phase1_synth5k.pdparams |
0.304 | 0.942 | Best model — 5,000 synthetic KiCad pre-training |
| v1 (exp6) | lora_exp6_best.pdparams |
0.119 | 0.946 | 1,500 real + synthetic text (20% mix) |
Benchmark on test_clean (N=30), greedy decoding, repetition_penalty=1.1, max_tokens=80.
Architecture
- Base model: PaddleOCR-VL-0.9B
- LoRA: r=16, alpha=32, target_modules=[q_proj, k_proj, v_proj, o_proj, linear_1, linear_2]
- Trainable params: 5.73M / 908M (0.63%)
- Training: single RTX 4060 8GB
Quick Start
# Load base model → apply LoRA → load weights via p.set_value()
# See github repo for full training/eval scripts.
Links
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support