72ea72b49c
- PEFT/LoRA fine-tuning (r=8, alpha=16) — only 1.2% params trainable - Data augmentation: templates + crypto slang (5x from 177→885 samples) - Human-in-the-loop CLI verification with persistent JSONL log - Disk-conscious: save_total_limit=1, ~6MB adapter size - CPU-friendly, single-file deployment - Tested: 1 epoch in ~4min on CPU, 6MB adapter
9 lines
222 B
Plaintext
9 lines
222 B
Plaintext
# Minimal requirements for LoRA pipeline (CPU-friendly)
|
|
torch==2.3.0 --index-url https://download.pytorch.org/whl/cpu
|
|
transformers==4.44.0
|
|
peft==0.12.0
|
|
datasets==2.20.0
|
|
accelerate==0.33.0
|
|
scikit-learn==1.5.0
|
|
numpy==1.26.0
|