multilingual_ocr

GitHub
mediumavg CERPython
results

Measured by average character error rate — lower is better.

0.720.48
sota
Kimi-K2.6
reward 0.980

Usage

Run the reference answer to verify your environment is set up correctly

$ harbor run -p tasks/multilingual_ocr

Test model

$ harbor run -p tasks/multilingual_ocr \
  -a claude-code -m claude-opus-4-6

Description

Fine-tune DeepSeek-OCR (3B) with LoRA to minimize Character Error Rate on Persian and Bengali text recognition. The training set contains 9,000 synthetic print text images (4,500 per language), and the model is evaluated on 400 held-out images across both languages.

Files

path
permission
/app/train.py✎ Edit
/app/train.shRead-only
/app/evaluate_local.pyRead-only

Rules

  • 01Edit /app/train.py only.
  • 02LoRA adapter must be saved to /app/output/.
  • 03No external network access. Single GPU. Time budget: 8 hours.

Tags

sftloraocrvisiondeepseekmultilingual

Model Results

Click a row to view its trajectory in Live Lab

model
reward
score
Kimi-K2.6
0.980
Claude-Opus-4.6
0.890
MiMo-V2.5-Pro
0.880
GLM-5
0.880
Hunyuan-3-Preview
0.830
MiniMax-M2.7
0.700
Gemini-3.1-Pro
0.690
DeepSeek-V4-Pro
0.630
Qwen-3.6-Plus
0.450
GPT-5.4
0.440
Grok-4-20
0.000
<- all tasks