abelrguezr/llm-instruction-finetuning

LLM Instruction Fine-Tuning

This skill guides you through fine-tuning a pre-trained LLM to follow instructions and respond to tasks like a chatbot, rather than just generating text.

When to Use This Skill

Use this skill when:

• You have a dataset of instruction-response pairs and want to fine-tune an LLM
• You need to format data in Alpaca or Phi-3 style for instruction tuning
• You want to evaluate a fine-tuned model's response quality
• You're building a chatbot or task-oriented AI assistant
• You need to understand the complete instruction fine-tuning workflow

Overview

Instruction fine-tuning transforms a pre-trained language model into one that:

• Understands it should respond to specific prompts
• Follows the format and style of instruction-response pairs
• Generates appropriate responses to user queries

Step 1: Prepare Your Dataset

Dataset Format

You need a dataset with instructions and responses. Common formats:

Alpaca Style:

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
Calculate the area of a circle with a radius of 5 units.

### Response:
The area of a circle is calculated using the formula A = πr². Plugging in the radius of 5 units:
A = π(5)² = π × 25 = 25π square units.

Phi-3 Style:

<|User|>
Can you explain what gravity is in simple terms?

<|Assistant|>
Absolutely! Gravity is a force that pulls objects toward each other.

Format Your Data

Use the format_instruction_data.py script (in scripts/) to convert your raw instruction-response pairs into the desired format. This handles:

• Adding instruction/response templates
• Handling optional input fields
• Creating training/validation/test splits

Step 2: Create Data Loaders

Key Processing Steps

1. Tokenize all texts using your tokenizer
2. Pad samples to the same length (typically your model's context length)
3. Create targets by shifting inputs by 1 token (for next-token prediction)
4. Mask padding tokens with -100 to exclude them from loss calculation
5. Optionally mask instruction tokens so the model only learns to generate responses

Why Mask with -100?

Using cross_entropy(..., ignore_index=-100) tells PyTorch to ignore targets with -100. This means:

• Padding tokens don't contribute to loss
• You can optionally mask the instruction portion so the model focuses on learning responses

Step 3: Load Pre-trained Model & Fine-Tune

Loading the Model

Load your pre-trained LLM (this was covered in previous training steps). Then use your existing training function to fine-tune.

Monitoring Training

Watch both training and validation loss:

• Training loss decreasing: Model is learning
• Validation loss decreasing: Model is generalizing
• Validation loss increasing while training loss decreases: Overfitting is occurring

Action: Stop training at the epoch where validation loss starts increasing to avoid overfitting.

Step 4: Evaluate Response Quality

Why Manual Evaluation Matters

Unlike classification tasks, you can't trust loss alone for instruction tuning. The model might:

• Generate correct format and syntax
• But give completely wrong answers

Loss won't catch this. You must evaluate response quality.

Evaluation Methods

1. Manual Review

• Generate responses on your test set
• Review them manually for correctness
• Check for hallucinations, wrong answers, or format issues

2. LLM-as-Judge

• Pass generated responses and expected responses to another LLM
• Ask it to evaluate quality, correctness, and helpfulness

3. Standardized Benchmarks

| Benchmark | What It Tests | Link | |-----------|---------------|------| | MMLU | Knowledge across 57 subjects (humanities, sciences, etc.) | https://arxiv.org/abs/2009.03300 | | LMSYS Chatbot Arena | Side-by-side chatbot comparison | https://arena.lmsys.org | | AlpacaEval | GPT-4 evaluates model responses | https://github.com/tatsu-lab/alpaca_eval | | GLUE | 9 NLU tasks (sentiment, entailment, QA) | https://gluebenchmark.com | | SuperGLUE | Harder version of GLUE | https://super.gluebenchmark.com | | BIG-bench | 200+ tasks (reasoning, translation, QA) | https://github.com/google/BIG-bench | | HELM | Comprehensive evaluation (accuracy, robustness, fairness) | https://crfm.stanford.edu/helm | | HumanEval | Code generation problems | https://github.com/openai/human-eval | | SQuAD | Question answering on Wikipedia | https://rajpurkar.github.io/SQuAD-explorer | | TriviaQA | Trivia questions with evidence | https://nlp.cs.washington.edu/triviaqa |

Common Issues & Solutions

Issue: Model Ignores Instructions

• Cause: Not enough instruction-response examples, or poor formatting
• Solution: Ensure consistent format, increase training data, verify masking is correct

Issue: Model Overfits

• Cause: Training too long, dataset too small
• Solution: Use early stopping on validation loss, add more diverse data

Issue: Model Generates Wrong Answers

• Cause: Model memorized format but not content
• Solution: Manual evaluation, use LLM-as-judge, check for hallucinations

Issue: Loss Doesn't Decrease

• Cause: Learning rate too high/low, data formatting issues
• Solution: Adjust learning rate, verify tokenization and masking

Best Practices

1. Start small: Fine-tune on a subset first to verify your pipeline works
2. Monitor both losses: Training AND validation loss tell different stories
3. Evaluate qualitatively: Don't trust loss alone for instruction tasks
4. Use early stopping: Prevent overfitting by watching validation loss
5. Test on diverse prompts: Ensure generalization beyond training distribution

References

• Build a Large Language Model from Scratch
• LLMs-from-scratch - Instruction Fine-Tuning Code