Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Parameter-Efficient Fine-Tuning (PEFT) library. LoRA, QLoRA, AdaLoRA, IA3, Prefix Tuning, P-Tuning, Prompt Tuning. Fine-tune large models with minimal memory overhead. Hugging Face ecosystem integration.
.claude/skills/mkurman-peft/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 3% | 0% |
| case-01 | ✓→✓ | ＝ Same ✓ | 16% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | 2% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | 179% | 0% |
| case-05 | ✓→✓ | ＝ Same ✓ | 52% | 0% |
PEFT (Parameter-Efficient Fine-Tuning) adapts large pretrained models by training only a small subset of parameters. Supports LoRA, QLoRA, AdaLoRA, IA3, Prefix Tuning, P-Tuning, and Prompt Tuning. Reduces GPU memory by 4-16x compared to full fine-tuning.
bashuv pip install peft
pythonfrom transformers import AutoModelForCausalLM from peft import LoraConfig, get_peft_model, TaskType model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct") peft_config = LoraConfig( r=16, lora_alpha=32, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], task_type=TaskType.CAUSAL_LM, ) model = get_peft_model(model, peft_config) model.print_trainable_parameters()
pythonmodel.save_pretrained("adapter") from peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct") merged = PeftModel.from_pretrained(base, "adapter").merge_and_unload()
pythonfrom transformers import BitsAndBytesConfig model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", quantization_config=BitsAndBytesConfig(load_in_4bit=True), device_map="auto") model = get_peft_model(model, peft_config)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 6,706 | 5,000 | -25% | 1 | 1 | 0% | 1,213 | 1,404 | +16% | 0 | 0 | — |
case-02 | pass→pass | 6,854 | 4,697 | -31% | 1 | 1 | 0% | 1,284 | 1,304 | +2% | 0 | 0 | — |
case-03 | fail→pass | 8,358 | 6,112 | -27% | 1 | 1 | 0% | 1,475 | 1,518 | +3% | 0 | 0 | — |
case-04 | pass→pass | 2,656 | 2,226 | -16% | 1 | 1 | 0% | 285 | 796 | +179% | 0 | 0 | — |
case-05 | pass→pass | 3,741 | 2,780 | -26% | 1 | 1 | 0% | 593 | 904 | +52% | 0 | 0 | — |
case-06 | pass→pass | 3,932 | 2,561 | -35% | 1 | 1 | 0% | 682 | 867 | +27% | 0 | 0 | — |
case-07 | pass→pass | 3,272 | 3,732 | +14% | 1 | 1 | 0% | 641 | 1,245 | +94% | 0 | 0 | — |
case-08 | pass→pass | 5,305 | 3,805 | -28% | 1 | 1 | 0% | 865 | 1,141 | +32% | 0 | 0 | — |
case-09 | pass→pass | 5,246 | 5,428 | +3% | 1 | 1 | 0% | 894 | 1,378 | +54% | 0 | 0 | — |
case-10 | pass→pass | 7,118 | 2,682 | -62% | 1 | 1 | 0% | 1,310 | 880 | -33% | 0 | 0 | — |
case-11 | pass→pass | 8,114 | 1,764 | -78% | 1 | 1 | 0% | 1,485 | 691 | -53% | 0 | 0 | — |
case-12 | pass→pass | 1,985 | 1,480 | -25% | 1 | 1 | 0% | 254 | 652 | +157% | 0 | 0 | — |
case-13 | pass→pass | 5,499 | 2,624 | -52% | 1 | 1 | 0% | 970 | 887 | -9% | 0 | 0 | — |
case-14 | pass→pass | 2,912 | 3,040 | +4% | 1 | 1 | 0% | 421 | 990 | +135% | 0 | 0 | — |
case-15 | pass→pass | 7,142 | 3,618 | -49% | 1 | 1 | 0% | 1,144 | 1,053 | -8% | 0 | 0 | — |
case-16 | pass→pass | 5,304 | 4,387 | -17% | 1 | 1 | 0% | 838 | 1,225 | +46% | 0 | 0 | — |
case-17 | pass→pass | 4,652 | 5,213 | +12% | 1 | 1 | 0% | 702 | 1,406 | +100% | 0 | 0 | — |
case-18 | pass→pass | 5,934 | 9,034 | +52% | 1 | 1 | 0% | 1,026 | 2,076 | +102% | 0 | 0 | — |
case-19 | pass→pass | 6,992 | 4,683 | -33% | 1 | 1 | 0% | 1,545 | 1,474 | -5% | 0 | 0 | — |
case-20 | pass→pass | 13,463 | 9,731 | -28% | 1 | 1 | 0% | 2,517 | 2,386 | -5% | 0 | 0 | — |
case-21 | pass→pass | 14,572 | 12,298 | -16% | 1 | 1 | 0% | 2,826 | 2,899 | +3% | 0 | 0 | — |
case-22 | pass→pass | 13,427 | 10,685 | -20% | 1 | 1 | 0% | 2,422 | 2,564 | +6% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +5 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.