Fine

Fine-tune machine learning models with Ruby.

classifier = Fine::TextClassifier.new("distilbert-base-uncased")
classifier.fit(train_file: "reviews.jsonl", epochs: 3)
classifier.predict("This product is amazing!")
# => [{ label: "positive", score: 0.97 }]

I want to fine-tune...

Text classification (sentiment, spam, intent)

Classify text into categories—reviews, support tickets, chat messages.

classifier = Fine::TextClassifier.new("distilbert-base-uncased")
classifier.fit(train_file: "data/reviews.jsonl", epochs: 3)

classifier.predict("Terrible experience, waste of money")
# => [{ label: "negative", score: 0.94 }]

Full tutorial: Text Classification

Text embeddings for semantic search

Train embeddings for your domain—legal docs, support tickets, product catalog.

embedder = Fine::TextEmbedder.new("sentence-transformers/all-MiniLM-L6-v2")
embedder.fit(train_file: "data/pairs.jsonl", epochs: 3)

embedding = embedder.encode("How do I cancel my subscription?")

Full tutorial: Text Embeddings

Image classification

Classify images into categories (cats vs dogs, products, documents).

classifier = Fine::ImageClassifier.new("google/siglip2-base-patch16-224")
classifier.fit(train_dir: "data/train", val_dir: "data/val", epochs: 3)

classifier.predict("photo.jpg")
# => [{ label: "cat", score: 0.95 }]

Full tutorial: Image Classification

Image recognition for custom objects

Teach the model to recognize your products, logos, or custom objects.

classifier = Fine::ImageClassifier.new("google/siglip2-base-patch16-384") do |config|
  config.epochs = 5
  config.learning_rate = 1e-4
end

classifier.fit(train_dir: "products/train")
classifier.save("models/product_detector")

Full tutorial: Object Recognition

Image similarity search

Find visually similar images in your catalog.

encoder = model.encoder
embedding = encoder.call(image_tensor)
similarity = cosine_similarity(embedding1, embedding2)

Full tutorial: Similarity Search

LLMs (with LoRA)

Fine-tune Gemma, Llama, Qwen and other open models using LoRA—train only 0.5% of parameters.

model = Fine::Models::CausalLM.from_pretrained("google/gemma-3-1b-it")

# Apply LoRA - only 0.5% of params trainable
Fine::LoRA.apply(model, rank: 32, alpha: 64)
#   Trainable params: 5.96M (0.46%)

# Train on your data
trainer = Fine::Training::LLMTrainer.new(model, config, train_dataset: dataset)
trainer.fit

# Merge weights and save
Fine::LoRA.merge!(model)
model.save("my_model")

Full tutorial: LLM Fine-tuning | LoRA Tool Calling

Installation

gem 'fine'

Requires Ruby 3.1+, LibTorch, and libvips.

Full installation guide | Quickstart

Quick setup (macOS):

brew install pytorch libvips
bundle install

Supported Models

Text Classification

Model	Parameters	Speed	Quality
`distilbert-base-uncased`	66M	Fast	Good
`bert-base-uncased`	110M	Medium	Better
`microsoft/deberta-v3-small`	44M	Fast	Great
`microsoft/deberta-v3-base`	86M	Medium	Best

Text Embeddings

Model	Parameters	Best For
`sentence-transformers/all-MiniLM-L6-v2`	22M	Fast, general purpose
`sentence-transformers/all-mpnet-base-v2`	110M	Better quality
`BAAI/bge-small-en-v1.5`	33M	Retrieval optimized
`BAAI/bge-base-en-v1.5`	110M	Best retrieval

Vision (SigLIP2)

Model	Parameters	Best For
`google/siglip2-base-patch16-224`	86M	Quick experiments
`google/siglip2-base-patch16-384`	86M	Good balance
`google/siglip2-large-patch16-256`	303M	Maximum accuracy
`google/siglip2-so400m-patch14-224`	400M	Best quality

LLMs

Model	Parameters	Best For
`google/gemma-3-1b-it`	1B	Fast experiments, tool calling
`meta-llama/Llama-3.2-1B`	1B	Fast experiments
`google/gemma-3-4b-it`	4B	Good balance
`Qwen/Qwen2-1.5B`	1.5B	Multilingual
`mistralai/Mistral-7B-v0.1`	7B	Best quality

Configuration

Fine::TextClassifier.new("distilbert-base-uncased") do |config|
  config.epochs = 3
  config.batch_size = 16
  config.learning_rate = 2e-5
  config.max_length = 256

  config.on_epoch_end do |epoch, metrics|
    puts "Epoch #{epoch}: #{metrics[:accuracy]}"
  end
end

Export for Deployment

Export fine-tuned models to production formats.

ONNX - For ONNX Runtime, TensorRT, OpenVINO:

classifier.export_onnx("model.onnx")
embedder.export_onnx("embedder.onnx")

GGUF - For llama.cpp, ollama (LLMs only):

llm.export_gguf("model.gguf", quantization: :q4_0)

Full tutorial: Model Export

Roadmap

SigLIP2 image classification
Text classification (BERT, DeBERTa)
Text embedding models
LLM fine-tuning (Gemma, Llama, Qwen)
ONNX & GGUF export
LoRA fine-tuning
QLoRA (4-bit quantized LoRA)

Contributing

Bug reports and pull requests welcome at github.com/khasinski/fine.

License

MIT

fine

Development

Runtime