Focus Patrol — Fine-tuned Qwen2.5-0.5B v7
Fine-tuned for 3-label browser tab classification: focus, neutral, distraction.
Test accuracy: 90% | Validation accuracy: 80%
Training
- Base model: Qwen/Qwen2.5-0.5B-Instruct
- Method: PEFT LoRA (r=16, alpha=32)
- Data: 156 labeled browsing examples (136 train + 20 validation)
- Epochs: 5 (early stopping at epoch 3)
- Hardware: Apple Silicon MPS
Usage with Transformers.js
import { pipeline } from '@huggingface/transformers';
const classifier = await pipeline('text-generation', 'rogeriobayer/focus-patrol-qwen2.5-0.5b-v7', {
dtype: 'q4',
device: 'webgpu',
});
const result = await classifier(prompt, {
max_new_tokens: 64,
temperature: 0.3,
do_sample: true,
return_full_text: false,
});
Performance
| Label | Test Accuracy | Validation Accuracy |
|---|---|---|
| Focus | 80% | 100% |
| Neutral | 100% | 73% |
| Distraction | 100% | 83% |
| Overall | 90% | 80% |
Training Report
See TRAINING_REPORT.md for full details.