Example Gallery

A fast index into the newer end-to-end examples: one entry per script, what pattern it demonstrates, and the concrete receipt it prints and asserts – a real measured number or invariant, not a restatement of the description. Each example is paired with a smoke test under mixle/tests/ that pins the same receipt programmatically.

For the full, general example inventory (distribution families, HMMs, enumeration, engines, and so on) see Examples. This page covers the newer applied adapter and multimodal-pretraining workflows, each landing as its own example.

Adapter and Multimodal Pretraining Patterns

Example	Pattern	Receipt
examples/peft_lora_grad_leaf.py	A real HuggingFace checkpoint (hf-internal-testing/tiny-random-gpt2), wrapped with actual peft.get_peft_model LoRA adapters, dropped into GradLeaf unchanged – next-token log-likelihood summed over a sequence is the log density GradLeaf expects.	Base checkpoint weights are bitwise unchanged after fitting (drift 0.0); only the LoRA adapter parameters move; mean sequence log-density rises from -48.27 to -46.18.
examples/multimodal_stage1_demo.py	LLaVA-style stage-1 pretraining on synthetic volumes: a frozen 3-D Conv3d encoder and a frozen toy LM (embedding + GRU cell + head) bridged by a thin trainable projection, fit end to end through one GradLeaf/optimize call.	All 11 frozen backbone tensors are bitwise unchanged (torch.equal before/after); the 4 projection tensors move; mean caption log-likelihood improves from -3.21 to -0.60.

Running the Examples

python examples/peft_lora_grad_leaf.py
python examples/multimodal_stage1_demo.py

peft_lora_grad_leaf.py additionally needs pip install "mixle[torch]" transformers peft (example-only dependencies, not a package extra, since GradLeaf has no opinion on what module it is handed). multimodal_stage1_demo.py needs only mixle[torch].

Every receipt above is pinned by a paired smoke test: mixle/tests/peft_lora_grad_leaf_smoke_test.py and mixle/tests/multimodal_stage1_demo_smoke_test.py.