Archived research examples

Archived examples preserve source, configuration, and provenance for historical study. Their commands require a separate historical checkout and environment; they are outside current runtime support. Read the per-family README before attempting restoration. Original documents inside an archive retain their old wording, including paths and commands that are no longer active.

Retired integrations and model-search variants

Family	Preserved scope
Legacy ViT factory	The former plato.models.vit factory and four ViT configs, including DeepViT, T2T-ViT, SwinV2, and LeViT.
FedTP	The complete FedTP experiment, including the server hypernetwork, algorithm, and both ViT/T2T-ViT configs.
pFedRLNAS / PerFedRLNAS	All NASViT, MobileNetV3, and DARTS modes, their shared configs, and vendored NASViT support.
FedRLNAS	The FedRLNAS experiment, its MNIST config, and its DARTS search-space code.
AnyCostFL local ViT	The local vit.py implementation and example_ViT.toml; shared runtime files are copied only as historical context.
FedRolex local ViT	The local vit.py implementation and example_ViT.toml; shared runtime files are copied only as historical context.
HeteroFL custom MobileNetV3	The custom mobilenetv3.py branch; shared runtime files and the ResNet config are copied only as historical context.
Nanochat	Original integration, runbook, and pinned upstream source snapshot.
LeRobot / SmolVLA	Original robotics integration and runbook; upstream policy/data revisions were not pinned.

The repository archive index links manifests, immutable source identities, historical restoration notes, and license provenance. Files marked copied_context explain a retired branch; they do not retire every current implementation with the same filename.

The active model-search examples retain AnyCostFL, FedRolex, and HeteroFL ResNet paths plus the separate SysHeteroFL ResNet example. The legacy model_type = "vit" factory is retired, while generic Hugging Face and Torchvision families remain separate. The current Hugging Face causal-LM factory is not an image-classification ViT replacement.

Qwen3 Federated LoRA is the maintained text-model reference. Its pinned base model and tokenizer do not make Nanochat artifacts interchangeable, and it does not replace robotics policy training. For Apple Silicon workloads, consult the separate native MLX guide.

Earlier historical examples and utilities

The four experiments below remain in examples/outdated/. Their 20 source and configuration files are listed individually so that an entrypoint, helper module or model definition is not mistaken for an independently supported example. The notes describe inspected source, not successful execution on current Python 3.13 dependencies. These files do not pin a complete working historical environment; dependency versions and the matching Plato checkout must be established separately before attempting a restoration.

Norm bounding

The server uses NumPy to calculate update norms and overrides FedAvg's aggregate_deltas hook. Its config selects Torchvision MNIST, LeNet-5 and SGD. The supplied threshold is 5, but omitting it leaves None in the denominator; threshold validation would need attention. Tensor-to-NumPy conversion also assumes tensors can be converted directly, so device handling needs explicit checks. The loop sums clipped updates without sample weighting or a final average. A restoration must check that aggregation rule against the cited research before changing it. The override hook still exists in current Plato; that alone does not qualify this implementation.

Preserved file	Role
norm_bounding.py	Experiment entrypoint.
norm_bounding_MNIST_lenet5.toml	MNIST/LeNet-5 configuration; threshold set to 5.
norm_bounding_server.py	Update norm calculation and clipped-update aggregation.

FjORD

The entrypoint imports both local model modules before selecting one. The preserved code uses PyTorch, NumPy, ptflops for complexity measurement and einops in its local ViT; even the ResNet entrypoint imports that ViT module. This is not evidence of a dependency on the separately retired legacy ViT factory. The supplied config selects CIFAR-10/ResNet-18 with stochastic width selection and resource limits disabled.

Restoration work would need to check model-factory/trainer interactions, subnet slicing, the distillation loss and aggregation of partially covered parameters. The optional resource-limit branch computes FLOPs but compares model size with both configured limits, which needs separate review. Existing client strategy code does not establish that these training and aggregation paths work together. The local ViT branch needs its own scope and validation; it is not covered by the ResNet planning estimate below.

Preserved file	Role
fjord.py	Entrypoint selecting a local ResNet or ViT model.
fjord_algorithm.py	Width selection, parameter slicing and aggregation.
fjord_client.py	Client lifecycle handling for the server-selected width.
fjord_resnet18_dynamic.toml	CIFAR-10/ResNet-18 configuration; resource limits disabled.
fjord_server.py	Per-client width responses and aggregation dispatch.
fjord_trainer.py	Server model construction and client subnet training.
resnet.py	Local variable-width ResNet model definitions.
vit.py	Local variable-width ViT model definition.

FL-MAML

This experiment uses PyTorch/NumPy, Torchvision MNIST and a custom trainer with separate inner and outer SGD optimizers, personalized testing and checkpoint paths. Several concrete source inconsistencies block a straightforward run: the trainer reads momentum and weight decay from Config().trainer, while the supplied config puts them under parameters.optimizer; scheduler branches use undefined optimizers and lr_scheduler names; and the two training_per_stage calls omit an argument required by its signature.

Fixing those interfaces would not establish MAML correctness. The connection between the copied inner-stage model and outer-stage gradients needs a numerical reference checked against the cited algorithm. Personalization isolation, report handling and checkpoint behavior also need end-to-end validation.

Preserved file	Role
fl_maml.py	Experiment entrypoint wiring client, server and trainer.
fl_maml_MNIST_lenet5.toml	MNIST/LeNet-5 configuration and meta learning rate.
fl_maml_client.py	Personalization request handling and accuracy reporting.
fl_maml_server.py	Training/personalization round and report coordination.
fl_maml_trainer.py	Two-stage training, personalization and testing loops.

CS-MAML

This experiment reuses the sibling FL-MAML client and trainer through sys.path.append("../fl_maml/"), making the import dependent on the working directory. It adds central/edge coordination, personalization events and accuracy reports on top of the same MNIST/PyTorch stack. It therefore inherits the FL-MAML trainer blockers above.

The entrypoint's four positional server.run arguments still match the current method signature; they are not, by themselves, a demonstrated API failure. Restoration would nevertheless need a real central/edge round to check model and trainer construction, report representation, event completion and personalization before claiming cross-silo compatibility.

Preserved file	Role
cs_maml.py	Cross-silo entrypoint importing sibling FL-MAML modules.
cs_maml_MNIST_lenet5.toml	MNIST/LeNet-5 configuration with one edge silo.
cs_maml_edge.py	Edge-client personalization strategy and event waiting.
cs_maml_server.py	Central/edge round, report and personalization coordination.

Restoration planning estimates

These are preliminary estimates in focused implementation/review sessions for bounded local regression fixtures. They are not commitments or promises of a working restoration. Recovering an appropriate dependency environment, resolving research semantics, obtaining external assets and reproducing paper-scale results can require additional work.

Experiment	Planning estimate	Work needed before claiming a restored example
Norm bounding	1–2 sessions	Threshold handling, a reference for the clipping/aggregation rule, and numerical/device regressions.
FjORD ResNet	3–5 sessions	Width selection, real subnet training and aggregation checks; local ViT restoration is separate.
FL-MAML	3–5 sessions	Trainer/config repairs, a two-stage gradient reference, personalization isolation and checkpoint qualification.
CS-MAML	2–3 additional sessions after FL-MAML	A real central/edge numerical round, personalization reporting and lifecycle checks.

Other preserved utilities

Multimodal dataset utilities and the legacy Gym adapter remain preserved utilities with their original limitations. They are separate from the four experiment inventories above; individual modules are not maintained experiment entrypoints. Consult their archive READMEs for original locations, dependencies and known limitations.

These are research references, not recommended current strategy templates. Start new client extensions from the active examples in the client reference. Preserved lockfiles and captured source hashes establish provenance, not successful execution on current Python 3.13 dependencies. External datasets, checkpoints, and toolchains may need separate historical versions; unknown upstream revisions remain unknown.