SKILL.md
Fine-Tuning Skill
Bonded to: prompt-optimization-agent
Quick Start
Skill("custom-plugin-prompt-engineering:fine-tuning")
Parameter Schema
parameters:
tuning_method:
type: enum
values: [full, lora, qlora, prompt_tuning, prefix_tuning]
default: lora
dataset_size:
type: enum
values: [small, medium, large]
description: "<1k, 1k-10k, >10k examples"
compute_budget:
type: enum
values: [low, medium, high]
default: medium
Tuning Methods Comparison
| Method |
Parameters |
Compute |
Quality |
Best For |
| Full Fine-tune |
All |
Very High |
Highest |
Maximum customization |
| LoRA |
~0.1% |
Low |
High |
Resource-constrained |
| QLoRA |
~0.1% |
Very Low |
Good |
Consumer GPUs |
| Prompt Tuning |
<0.01% |
Minimal |
Good |
Simple tasks |
| Prefix Tuning |
~0.1% |
Low |
Good |
Generation tasks |
Dataset Preparation
Format Templates
formats:
instruction:
template: |
### Instruction
{instruction}
### Response
{response}
chat:
template: |
<|user|>
{user_message}
<|assistant|>
{assistant_response}
completion:
template: "{input}{output}"
Quality Criteria
quality_checklist:
- [ ] No duplicate examples
- [ ] Consistent formatting
- [ ] Diverse examples
- [ ] Balanced categories
- [ ] High-quality outputs
- [ ] No harmful content
Training Configuration
training_config:
hyperparameters:
learning_rate: 2e-5
batch_size: 8
epochs: 3
warmup_ratio: 0.1
lora_config:
r: 16
alpha: 32
dropout: 0.05
target_modules: ["q_proj", "v_proj"]
evaluation:
eval_steps: 100
save_steps: 500
metric: loss
Evaluation Framework
| Metric |
Purpose |
Target |
| Loss |
Training progress |
Decreasing |
| Accuracy |
Task performance |
>90% |
| Perplexity |
Model confidence |
<10 |
| Human eval |
Quality assessment |
Preferred >80% |
Troubleshooting
| Issue |
Cause |
Solution |
| Overfitting |
Small dataset |
Add regularization |
| Underfitting |
Low epochs |
Increase training |
| Catastrophic forgetting |
Aggressive tuning |
Lower learning rate |
| Poor generalization |
Data bias |
Diversify dataset |
References
See: Hugging Face PEFT, OpenAI Fine-tuning Guide
