Fine-tune LLMs in 3 lines.

Smart defaults

Learning rate, batch size and LoRA adapter size start from values that suit your model, method and GPU. Change any of them; none is required.

VRAM-aware training

The adapter size and batch size are chosen to fit the memory free on your GPU, and an estimate says whether a run fits before it starts. Works from 8 GB cards (1B to 3B models) to 32 GB (up to 32B with QLoRA).

First-class Windows

Developed on Windows with CUDA. It handles the usual PyTorch and Unsloth pitfalls there for you. One feature, full fine-tuning with offload, needs Linux or WSL2.

Install

# Recommended
pip install backpropagate[standard]

# Minimal core only
pip install backpropagate

# All extras
pip install backpropagate[full]

# Requires: Python 3.10+ · CUDA GPU (8GB+ VRAM)

Basic training

from backpropagate import Trainer

# Smart defaults — no config needed
trainer = Trainer("unsloth/Qwen2.5-7B-Instruct-bnb-4bit")
trainer.train("my_data.jsonl", steps=100)
trainer.save("./my-model")

Multi-run SLAO

from backpropagate.multi_run import MultiRunTrainer

runner = MultiRunTrainer(
    model="unsloth/Llama-3.2-3B-Instruct-bnb-4bit",
    num_runs=5, steps_per_run=100,
    merge_mode="slao",
)
result = runner.run("my_data.jsonl")

Export to Ollama

from backpropagate.export import export_gguf, register_with_ollama

result = export_gguf(model, tokenizer, "./output", quantization="q4_k_m")
register_with_ollama(result.path, model_name="my-model")
# ollama run my-model

Headless by design

No UI required. Runs in CI/CD pipelines, SSH sessions, and automated workflows. Full Python API with structured logging. Callbacks for progress tracking and early stopping.

Multi-run SLAO

Single LoRA Continual Learning via Asymmetric Merging (arXiv:2512.23017) prevents catastrophic forgetting during extended fine-tuning campaigns via orthogonal init, asymmetric A/B handling, and time-aware scaling. Checkpoint-and-resume keeps long runs recoverable after crashes.

LoRA + QLoRA + full FT + Unsloth

LoRA, QLoRA (4-bit) and full fine-tuning: up to about 6B on a 32 GB card, and a 7B-class model with offload on Linux or WSL2. Unsloth-accelerated when installed. Export to GGUF at q2_k, q4_k_m, q8_0 or f16.