Examples¶

Practical examples for common fine-tuning scenarios.

Example 1: Fine-tune GPT-2 on Custom Dialogue Data¶

Scenario¶

You have a JSONL file with conversational data and want to create a chatbot.

Dataset Format (dialogue.jsonl)¶

{"prompt": "Hello, how are you?", "response": "I'm doing great! How can I help you today?"}
{"prompt": "What's the weather like?", "response": "I don't have access to weather data, but you can check weather.com"}

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: gpt2
Output directory: ./dialogue_model
Dataset path: ./dialogue.jsonl
Limit samples: yes
Number of samples: 5000
Max sequence length: 256
LoRA r: 8
LoRA alpha: 32
LoRA dropout: 0.1
Epochs: 5
Batch size: 8
Learning rate: 2e-4
Upload to HuggingFace: no

Expected Results¶

📊 PERFORMANCE COMPARISON
Metric       Base Model      Fine-tuned      Improvement
ROUGE1       0.1823          0.3245          +78.03%
ROUGE2       0.0912          0.2134          +133.99%
ROUGEL       0.1645          0.2987          +81.58%

Using the Fine-tuned Model¶

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("gpt2")
tokenizer = AutoTokenizer.from_pretrained("gpt2")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "./dialogue_model")

# Generate response
prompt = "Hello, how are you?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Example 2: Domain Adaptation for Medical Text¶

Scenario¶

Adapt a model to medical terminology and clinical notes.

Dataset¶

Using HuggingFace medical dataset.

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: facebook/opt-350m
Output directory: ./medical_model
Dataset name: medical_meadow_medical_flashcards
Dataset config: default
Split: train
Limit samples: yes
Number of samples: 20000
Max sequence length: 512
LoRA r: 16
LoRA alpha: 64
LoRA dropout: 0.15
Epochs: 3
Batch size: 4
Learning rate: 1e-4
Upload to HuggingFace: yes
Repo name: myusername/opt-medical-350m

Why These Settings?¶

• Higher rank (16): Medical domain requires learning specialized terminology
• Higher alpha (64): Stronger adaptation to domain-specific patterns
• More dropout (0.15): Medical text can be noisy, prevent overfitting
• Lower learning rate (1e-4): Conservative to preserve general knowledge

Example 3: Summarization Task¶

Scenario¶

Fine-tune for news article summarization.

Dataset Format (summaries.csv)¶

article,summary
"Long article text here...","Brief summary here..."
"Another article...","Another summary..."

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: google/flan-t5-base
Output directory: ./summarization_model
Dataset path: ./summaries.csv
Limit samples: yes
Number of samples: 15000
Max sequence length: 1024
LoRA r: 8
LoRA alpha: 32
LoRA dropout: 0.1
Epochs: 3
Batch size: 4
Learning rate: 2e-4

Data Preparation Tips¶

The tool auto-detects columns. For best results:

1. Name columns clearly: article, text, summary, content
2. Clean your data: Remove HTML tags, special characters
3. Balance length: Keep articles similar length when possible

Example 4: Code Generation¶

Scenario¶

Fine-tune on code examples for Python code generation.

Dataset¶

Using HuggingFace code dataset with specific file selection.

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: Salesforce/codegen-350M-mono
Output directory: ./code_model
Dataset name: codeparrot/github-code
Dataset config: python
Load specific file: yes
File path: train-00000-of-00200.parquet
Number of samples: 10000
Max sequence length: 512
LoRA r: 8
LoRA alpha: 32
LoRA dropout: 0.05
Epochs: 2
Batch size: 8
Learning rate: 2e-4

Why These Settings?¶

• Lower dropout (0.05): Code has consistent structure, less noise
• Fewer epochs (2): Code datasets are large, less overfitting risk
• Specific file selection: Avoids loading entire 200-file dataset

Example 5: Question Answering¶

Scenario¶

Fine-tune on SQuAD-style Q&A data.

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: distilbert-base-uncased
Output directory: ./qa_model
Dataset name: squad
Dataset config: plain_text
Split: train
Limit samples: yes
Number of samples: 30000
Max sequence length: 384
LoRA r: 8
LoRA alpha: 32
LoRA dropout: 0.1
Epochs: 3
Batch size: 16
Learning rate: 3e-4

Inference Example¶

from transformers import pipeline
from peft import PeftModel, AutoModelForQuestionAnswering

# Load fine-tuned model
base_model = AutoModelForQuestionAnswering.from_pretrained("distilbert-base-uncased")
model = PeftModel.from_pretrained(base_model, "./qa_model")

# Create QA pipeline
qa = pipeline("question-answering", model=model, tokenizer="distilbert-base-uncased")

# Ask question
context = "Paris is the capital of France. It is known for the Eiffel Tower."
question = "What is the capital of France?"
result = qa(question=question, context=context)
print(result['answer'])  # "Paris"

Example 6: Multi-language Fine-tuning¶

Scenario¶

Fine-tune multilingual model for specific languages.

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: xlm-roberta-base
Output directory: ./multilingual_model
Dataset name: Helsinki-NLP/tatoeba
Dataset config: eng-spa
Split: train
Limit samples: yes
Number of samples: 25000
Max sequence length: 128
LoRA r: 16
LoRA alpha: 32
LoRA dropout: 0.1
Epochs: 4
Batch size: 8
Learning rate: 2e-4

Example 7: Sentiment Analysis¶

Scenario¶

Adapt model for sentiment classification.

Dataset Format (sentiment.jsonl)¶

{"text": "This product is amazing!", "sentiment": "positive"}
{"text": "Terrible experience, would not recommend.", "sentiment": "negative"}

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: roberta-base
Output directory: ./sentiment_model
Dataset path: ./sentiment.jsonl
Limit samples: yes
Number of samples: 10000
Max sequence length: 256
LoRA r: 4
LoRA alpha: 16
LoRA dropout: 0.1
Epochs: 5
Batch size: 16
Learning rate: 3e-4

Why These Settings?¶

• Lower rank (4): Sentiment is relatively simple classification
• Higher batch size (16): Shorter sequences allow larger batches
• More epochs (5): Smaller dataset benefits from more iterations

Example 8: Instruction Following¶

Scenario¶

Fine-tune on instruction-response pairs.

Dataset¶

Using large instruction dataset with selective loading.

Configuration¶

python xlmtec.py

# Interactive prompts:
Model name: EleutherAI/pythia-410m
Output directory: ./instruction_model
Dataset name: HuggingFaceH4/ultrachat_200k
Load specific file: yes
File path: data/train_sft-00000-of-00004.parquet
Number of samples: 15000
Max sequence length: 1024
LoRA r: 16
LoRA alpha: 64
LoRA dropout: 0.1
Epochs: 2
Batch size: 4
Learning rate: 1e-4
Upload to HuggingFace: yes
Repo name: myusername/pythia-instruction-410m
Private: no

Best Practices from Examples¶

Dataset Size Guidelines¶

• Experiments: 1,000-5,000 samples
• Development: 5,000-20,000 samples
• Production: 20,000+ samples

Model Selection Tips¶

1. Start small: Test with GPT-2 or OPT-125M
2. Match architecture: Use decoder models (GPT) for generation
3. Consider license: Check model licensing for commercial use
4. Resource awareness: Larger models need more VRAM

Common Pitfalls to Avoid¶

1. ❌ Training on too few samples (< 500)
2. ❌ Using max_length longer than needed (wastes memory)
3. ❌ Setting rank too high for simple tasks (overfitting)
4. ❌ Forgetting to validate results before uploading
5. ❌ Not monitoring GPU memory usage

Performance Optimization¶

# For faster iteration during development:
- Use smaller model variants
- Limit samples to 1000-5000
- Reduce max_length
- Use rank 4-8

# For production quality:
- Use full dataset
- Increase rank to 16
- Train for more epochs
- Validate on held-out test set

Recipes Summary¶

Next Steps¶

• Learn about Configuration Options
• Check Troubleshooting for common issues
• Review API Reference for programmatic usage