日本語 | 中文 | Español | Français | हिन्दी | Italiano | Português (BR)

Headless LLM fine-tuning in 3 lines. Smart defaults, VRAM-aware batch sizing, multi-run SLAO, and one-click GGUF export for Ollama.

Train LLMs in 3 lines of code. Export to Ollama in one more.

Quick Start

pip install backpropagate[standard]

from backpropagate import Trainer

trainer = Trainer("unsloth/Qwen2.5-7B-Instruct-bnb-4bit")
trainer.train("my_data.jsonl", steps=100)
trainer.export("gguf", quantization="q4_k_m")  # Ready for Ollama

Dataset Format

Your JSONL training file should have one example per line. The simplest format is ShareGPT chat:

{"conversations": [{"from": "human", "value": "What is Python?"}, {"from": "gpt", "value": "A programming language."}]}
{"conversations": [{"from": "human", "value": "Explain recursion."}, {"from": "gpt", "value": "A function that calls itself."}]}

Alpaca (instruction/output), OpenAI chat (messages), and raw text formats are also supported.

Why Backpropagate?

Problem	Solution
Fine-tuning is complex	3 lines: load, train, save
Windows is a nightmare	First-class Windows support
VRAM management is hard	Auto batch sizing, GPU monitoring
Model export is confusing	One-click GGUF + Ollama registration
Long runs cause forgetting	Multi-run SLAO training

Key Features

• Headless by Design: Built for CI/CD pipelines, automated workflows, and programmatic execution.
• Smart Defaults: Automatically configures optimal hyperparameters based on your hardware and dataset.
• Multi-Run SLAO Training: Advanced training strategies to prevent catastrophic forgetting during long runs.
• First-Class Windows Support: Tested and optimized for Windows environments, avoiding common PyTorch/CUDA pitfalls.
• Seamless Export: One-click export to GGUF format and automatic registration with Ollama.
• Modular Architecture: Install only the dependencies you need (e.g., [unsloth], [ui], [export]).

Installation

pip install backpropagate              # Core only (minimal)
pip install backpropagate[unsloth]     # + Unsloth 2x faster training
pip install backpropagate[ui]          # + Gradio web UI
pip install backpropagate[standard]    # unsloth + ui (recommended)
pip install backpropagate[full]        # Everything

Extra	Description	Dependencies
unsloth	2x faster training, 50% less VRAM	unsloth
ui	Gradio web interface	gradio>=5.6.0
validation	Pydantic config validation	pydantic, pydantic-settings
export	GGUF export for Ollama	llama-cpp-python
monitoring	WandB + system monitoring	wandb, psutil
observability	OpenTelemetry tracing	opentelemetry-api, opentelemetry-sdk
logging	Structured logging	structlog
security	JWT auth + token generation	PyJWT, cryptography
production	unsloth + ui + validation + logging + security	(bundle)

Requirements: Python 3.10+ · CUDA GPU (8GB+ VRAM) · PyTorch 2.0+

Configuration

All settings can be overridden with environment variables using the BACKPROPAGATE_ prefix (e.g., BACKPROPAGATE_LOG_LEVEL=debug). A .env file in the project root is loaded automatically when the [validation] extra is installed.

Usage

Basic Training

from backpropagate import Trainer

trainer = Trainer("unsloth/Qwen2.5-7B-Instruct-bnb-4bit")
trainer.train("my_data.jsonl", steps=100)
trainer.save("./my-model")
trainer.export("gguf", quantization="q4_k_m")

Multi-Run SLAO Training

from backpropagate import Trainer

trainer = Trainer("unsloth/Qwen2.5-7B-Instruct-bnb-4bit")

result = trainer.multi_run(
    dataset="HuggingFaceH4/ultrachat_200k",
    num_runs=5,
    steps_per_run=100,
    samples_per_run=1000,
    merge_mode="slao",  # Smart LoRA merging
)

Export to Ollama

# Export to GGUF
result = trainer.export("gguf", quantization="q4_k_m")

# Register with Ollama separately
from backpropagate import register_with_ollama
register_with_ollama(result.path, "my-finetuned-model")
# ollama run my-finetuned-model

CLI

backprop train --data my_data.jsonl --model unsloth/Qwen2.5-7B-Instruct-bnb-4bit --steps 100
backprop multi-run --data my_data.jsonl --runs 5 --steps 100
backprop export ./output/lora --format gguf --quantization q4_k_m --ollama --ollama-name my-model
backprop ui --port 7862
backprop info

Windows Support

Backpropagate is designed to work on Windows out of the box:

• Pre-tokenization to avoid multiprocessing crashes
• Automatic xformers disable for RTX 40/50 series
• Safe dataloader settings
• Tested on RTX 5080 (16GB VRAM)

Model Presets

Preset	VRAM	Speed	Quality
Qwen 2.5 7B	~12GB	Medium	Best
Qwen 2.5 3B	~8GB	Fast	Good
Llama 3.2 3B	~8GB	Fast	Good
Llama 3.2 1B	~6GB	Fastest	Basic
Mistral 7B	~12GB	Medium	Good

Architecture

backpropagate/
├── trainer.py           # Core Trainer class
├── multi_run.py         # Multi-run SLAO training
├── slao.py              # SLAO LoRA merging algorithm
├── datasets.py          # Dataset loading, filtering & curriculum
├── export.py            # GGUF/Ollama export
├── config.py            # Pydantic settings + training presets
├── gpu_safety.py        # GPU monitoring & safety
├── cli.py               # CLI entry point (backprop command)
├── checkpoints.py       # Checkpoint management
├── exceptions.py        # Structured error hierarchy
├── feature_flags.py     # Optional feature detection
├── security.py          # Path traversal & torch security
├── logging_config.py    # Structured logging setup
├── theme.py             # Gradio theme customization
├── ui.py                # Gradio interface
└── ui_security.py       # Rate limiting, CSRF, file validation

Privacy

All training happens locally on your GPU. Backpropagate makes no network requests except to download models from HuggingFace (which you initiate). No telemetry, no cloud dependency.

Scorecard

Category	Score	Notes
A. Security	6/8	SECURITY.md, trust model, no secrets/telemetry, safe_path(). MCP items skipped
B. Error Handling	3/7	Structured exceptions + exit codes + no raw stacks. MCP/desktop/vscode skipped
C. Operator Docs	4/7	README, CHANGELOG, LICENSE, --help. Logging/MCP/complex skipped
D. Shipping Hygiene	6/9	verify.sh, version=tag, 5 scanners in CI, dependabot, python_requires, clean build
E. Identity	4/4	Logo, translations, landing page, metadata
Total	23/31	14 items skipped with justification · shipcheck audit passes 100%

License

MIT — see LICENSE for details.

---

Built by MCP Tool Shop