sage-finetune

Fine-tuning component implementations for SAGE L3.

Installation

pip install isage-finetune

For LoRA training:

pip install isage-finetune[peft]

Features

• LoRA Trainer: Parameter-efficient fine-tuning with Low-Rank Adaptation
• Mock Trainer: Testing trainer for pipeline validation
• JSON/JSONL Loader: Flexible data loading for instruction and chat formats

Boundary

• FinetuneManager only manages task scheduling and status lifecycle.
• Trainers only implement training/evaluation behavior.
• Package registration is explicit and fail-fast.

Quick Start

from sage_libs.sage_finetune import MockTrainer, JSONDatasetLoader

# Load training data
loader = JSONDatasetLoader()
train_data = loader.load("train.jsonl")

# Train (mock for testing)
trainer = MockTrainer()
result = trainer.train(train_data)
print(f"Loss: {result['train_loss']:.4f}")

LoRA Fine-tuning

from sage_libs.sage_finetune import LoRATrainer
from sage_libs.sage_finetune.trainers.lora_trainer import LoRAConfig

trainer = LoRATrainer(
    model_name="gpt2",
    lora_config=LoRAConfig(r=8, lora_alpha=16),
)

result = trainer.train(train_dataset)
trainer.save_model("./my_lora_model")

Data Formats

Instruction Format

{"instruction": "Summarize this text", "input": "Long text...", "output": "Summary..."}

Chat Format

{"messages": [{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Hi"}]}

Integration with SAGE

Components register into the SAGE finetune factory on package import:

from sage.libs.finetune import create_trainer

trainer = create_trainer("lora", model_name="gpt2")

License

Apache 2.0