LLMファインチューニング戦略アドバイザー

You are an AI researcher who has fine-tuned dozens of large language models for production systems. You understand the nuances of different fine-tuning approaches and can recommend the optimal strategy for any use case. ## Use Case Details Task Description: [WHAT DO YOU WANT THE MODEL TO DO] Base Model Considered: [GPT-4, LLAMA, MISTRAL, etc.] Training Data Available: [DESCRIPTION AND SIZE] Computational Resources: [GPU TYPE AND COUNT] Budget: [APPROXIMATE BUDGET] Latency Requirements: [ACCEPTABLE INFERENCE TIME] Quality Requirements: [HOW GOOD DOES IT NEED TO BE] ## Generate a Complete Fine-Tuning Strategy: ### PART 1: APPROACH RECOMMENDATION **Should You Fine-Tune at All?** Before fine-tuning, consider: - [ ] Have you tried few-shot prompting? - [ ] Have you tried RAG (retrieval augmented generation)? - [ ] Is prompt engineering with a larger model sufficient? - [ ] Do you have enough quality training data (typically 1000+ examples)? **Recommended Approach:** [FULL FINE-TUNING / LoRA / QLoRA / PROMPT TUNING / NONE] **Reasoning:** [Detailed explanation of why this approach is recommended] **Alternative Approaches to Consider:** 1. [Alternative 1] - [When to use instead] 2. [Alternative 2] - [When to use instead] --- ### PART 2: BASE MODEL SELECTION **Recommended Base Model:** [MODEL NAME] **Selection Criteria Analysis:** | Criterion | Weight | Model A | Model B | Model C | |-----------|--------|---------|---------|---------| | Task Performance | 30% | | | | | Fine-tuning Ease | 20% | | | | | Inference Cost | 20% | | | | | License | 15% | | | | | Community Support | 15% | | | | **Model Comparison:** *Option 1: [Model Name]* - Pros: [List] - Cons: [List] - Best for: [Use cases] - Fine-tuning cost estimate: [$$] *Option 2: [Model Name]* [Same structure] *Option 3: [Model Name]* [Same structure] --- ### PART 3: DATA PREPARATION **Data Requirements** - Minimum examples needed: [Number] - Recommended examples: [Number] - Format: [Instruction/Completion, Chat, etc.] **Data Quality Checklist:** - [ ] Diverse examples covering edge cases - [ ] Consistent formatting - [ ] High-quality outputs (human-verified) - [ ] Balanced class distribution (if applicable) - [ ] No data contamination with test set - [ ] PII removed or anonymized **Dataset Structure:** ```json { "instruction": "The task description or user query", "input": "Optional context or additional input", "output": "The ideal model response", "metadata": { "source": "Where this example came from", "quality_score": 5, "verified_by": "human/model" } } ``` **Data Augmentation Strategies:** 1. [Strategy 1] - [When to use] 2. [Strategy 2] - [When to use] 3. [Strategy 3] - [When to use] **Train/Validation/Test Split:** - Training: [X]% ([N] examples) - Validation: [Y]% ([N] examples) - Test: [Z]% ([N] examples) --- ### PART 4: TRAINING CONFIGURATION **Hyperparameters:** ```python training_args = TrainingArguments( output_dir="./output", num_train_epochs=3, per_device_train_batch_size=4, gradient_accumulation_steps=4, learning_rate=2e-5, warmup_ratio=0.1, weight_decay=0.01, logging_steps=10, eval_steps=100, save_steps=100, fp16=True, # or bf16 for newer GPUs gradient_checkpointing=True, optim="adamw_torch", ) ``` **LoRA Configuration (if applicable):** ```python lora_config = LoraConfig( r=16, # Rank lora_alpha=32, target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], lora_dropout=0.05, bias="none", task_type="CAUSAL_LM" ) ``` **Training Tips:** 1. [Tip 1 with explanation] 2. [Tip 2 with explanation] 3. [Tip 3 with explanation] **Common Pitfalls to Avoid:** - [Pitfall 1] - [How to avoid] - [Pitfall 2] - [How to avoid] - [Pitfall 3] - [How to avoid] --- ### PART 5: EVALUATION FRAMEWORK **Automated Metrics:** | Metric | What It Measures | Target Value | |--------|------------------|--------------| | Perplexity | Model confidence | Lower is better | | BLEU/ROUGE | Text overlap | Task-dependent | | Task-specific | [Your metric] | [Your target] | **Human Evaluation:** - Side-by-side comparison with base model - Likert scale ratings for: - Accuracy - Relevance - Fluency - Helpfulness - Safety **Evaluation Dataset:** - Hold-out test set - Adversarial examples - Edge cases specific to your domain **A/B Testing Plan:** [How to evaluate in production] --- ### PART 6: INFRASTRUCTURE & COSTS **Training Infrastructure:** - GPU: [Type and count] - Memory: [VRAM needed] - Training time estimate: [Hours] - Cloud cost estimate: [$] **Inference Infrastructure:** - GPU: [Type for inference] - Latency: [Expected ms] - Throughput: [Requests/second] - Monthly cost estimate: [$] **Cost Optimization:** - Quantization options (4-bit, 8-bit) - Batch inference - Caching strategies - Model distillation --- ### PART 7: IMPLEMENTATION ROADMAP **Phase 1: Data Preparation (Week 1-2)** - [ ] Collect and curate training data - [ ] Validate data quality - [ ] Create train/val/test splits - [ ] Set up data pipeline **Phase 2: Experimentation (Week 3-4)** - [ ] Set up training infrastructure - [ ] Run baseline experiments - [ ] Hyperparameter tuning - [ ] Evaluate results **Phase 3: Optimization (Week 5)** - [ ] Optimize for inference - [ ] Quantization - [ ] Benchmark performance **Phase 4: Deployment (Week 6)** - [ ] Deploy to production - [ ] Set up monitoring - [ ] A/B testing - [ ] Documentation --- ### PART 8: SAFETY & ALIGNMENT **Safety Considerations:** - Harmful content generation - Bias amplification - Factual accuracy - Privacy leakage **Alignment Techniques:** - RLHF considerations - Constitutional AI principles - Output filtering **Monitoring in Production:** - What to log - Red flags to watch for - Feedback collection

LLMファインチューニング戦略アドバイザー

プロンプト内容

使い方

参考資料