Version 0.2.0

[martinkorelic]

MobileTransformers v0.2.0

[0.2.0] — 17-08-2026

Added

Host — exporting models

• One-command export: a Hugging Face checkpoint becomes an ONNX inference graph, a PEFT-enabled
  training graph with an optimiser, a tokenizer, an optional embedding stage, and a manifest tying
  them together.
• PEFT methods: LoRA, LoRA-XS, and MARS (Multi-Adapter Rank Sharing) — this project's own
  method, which shares adapter components across layers so trainable parameters grow with rank
  rather than with depth.
• Decoder and encoder tasks: text generation, sequence classification, and feature extraction for
  retrieval.
• A mobiletransformers CLI covering export, packaging, validation, Hub push/pull, adapter
  conversion and federated aggregation.

Android SDK

• mobiletransformers-android — an AAR consumable from any app, with a single fromPretrained
  entry point that resolves, downloads, verifies and atomically installs a package.
• Generation — streaming, with a chat template and KV cache, over either of two selectable ONNX
  Runtime engines.
• Fine-tuning — a real training loop with an optimiser and a live loss curve, surviving the app
  going to the background.
• Merging — folding a trained adapter back into the inference weights, so the fine-tuned model is
  the one that then generates.
• Retrieval — chunking, embedding and vector search over documents you supply, with grounded
  generation that reports its sources before the answer.
• Classification for encoder packages, scored per label.
• Tool calling — a model's answer parsed into a structured call, validated against an allowlist
  the app owns, and bound to an Android intent. Nothing executes without user consent.
• Federated adapter exchange — export local factors, aggregate on a host, import the average
  back. Default-off and consent-gated.

All of it runs on the device, with no server and no data leaving the phone.

Published artifacts

• Six model packages on the Hub under
  mobiletransformers, each shipping both an inference
  and a training stage — including one exported with MARS.
• The gitignored native build artifacts are hosted, so a fresh clone can provision itself with
  make fetch-native-deps and no credentials.

Sample app and documentation

• A reference app exercising every capability through the public facade only, with navigation that
  derives from what the loaded package can actually do.
• A documentation site built from this repository's own docs/, and make doctor — a preflight
  report naming every missing prerequisite and the command that fixes it.