Standard libraries first

toolbox is a place to combine the best tools, not to rewrite them. If a well-maintained library already does something, toolbox uses it as a backend behind the same interfaces as everything else. Custom code is written only when there's a reason: to teach how something works, to change its internals, or because nothing standard exists.

That gives each area two layers:

• Standard backend: the library people use in production and papers. Use it by default.
• Built-in reference: a short, readable implementation for teaching and experimenting. Where both exist, the built-in is tested against the standard one.

Area	Standard backend in toolbox	Built-in reference	Select with
Language models, tokenizers, LoRA	Hugging Face transformers, peft	tiny_gpt, byte tokenizer	model=qwen2_5_0_5b_lora, tokenizer=qwen2_5
Post-training (SFT, DPO, GRPO, reward models)	TRL	supervised, preference, grpo trainers	trainer.name=trl trainer.algorithm=grpo
GNN layers and models	PyTorch Geometric	gcn, sage, gat (parity-tested against PyG)	model.conv=pyg_gcn, model.name=pyg_model
Your own model	any nn.Module	—	model: {name: torch_module, target: pkg.mod:Class}
Optimizers, LR schedules	torch.optim, transformers schedules	adamw, cosine ...	optimizer: {name: torch, optimizer: RAdam}, scheduler: {name: hf, schedule: cosine}
Datasets	Hugging Face datasets	JSONL, builtin:// samples	data.path=hf://org/name?split=train
Decision models	AnyJev (readout, evaluation)	toolbox.decision	Jev pipeline
Serving	SGLang, vLLM	torch engine	toolbox serve --engine sglang
RL environments	Gymnasium	cartpole	env=gymnasium env.env_id=...
Embeddings for memory	sentence-transformers	hashing embedder	SentenceTransformerEmbedder(...)
Configs, tracking	Hydra, Weights & Biases, TensorBoard	console, JSONL	callbacks

Each backend is an optional extra (pip install 'toolbox[llm]', [gnn], [jev], [serve-vllm], [embeddings] ...), imported only when used. If an extra is missing, the error names it.

Planned backends

Areas where a standard exists but toolbox still uses its own code are tracked on the roadmap: Accelerate / DeepSpeed for multi-GPU, OpenTelemetry for agent traces, and lm-evaluation-harness for evaluation.

Adding a backend

A backend is an adapter: a registered component that translates between the library's API and the toolbox interface for that role. Examples to copy:

• a layer: src/toolbox/nn/layers/pyg.py
• a model: src/toolbox/nn/models/external.py
• a trainer: src/toolbox/trainer/trl_backend.py
• an inference engine: src/toolbox/serving/engines.py