#3030th of 41 in Training and fine-tuning
fractalworkspace
Federated training and split-model inference across Android phones
- Stars
- 8.7k
- License
- custom license
- Last commit
- Oct 2026
- Language
- Python
535 on 15 May 20268,738 on 30 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Fractal pairs a Python Flask control plane (FractalCore) with a Kotlin Android node app. The server splits datasets into segments, aggregates checkpoint deltas with FedAvg, and slices models such as Llama 3 8B or TinyLlama into INT4 .pte partitions of 150MB or less that phones run as a pipeline over Wi-Fi. Nodes pause when battery is under 50%, not charging, or battery temperature exceeds 40 C.
Who it is for: Researchers and engineers experimenting with edge federated learning on Android
Strengths
- Covers both on-device federated training and layer-sliced inference across phones
- Raw datasets stay on devices; only checkpoint deltas go to the server
- Explicit limits: 150MB partitions, 4,100-byte activation frames, single-use task_Id
- Per-tenant storage isolation and X-Auth-Token auth on tenant and admin routes
Weaknesses
- Proprietary non-commercial license; commercial use needs written authorization
- Requires a Firebase project with Firestore for the token and reward ledger
- Needs a physical Android device (API 24+); no iOS or desktop nodes documented
- No release published; README gives no benchmarks or latency figures
What it needs
- Needs Firebase Firestore, Android Studio, Python 3.10+, Docker Compose, TFLite, ExecuTorch, torchao, XNNPACK
- Models: Llama 3 8B, TinyLlama, TFLite (.tflite), ExecuTorch (.pte)
- port 5000
Quick answers
Is fractalworkspace open source?
Its code is public on GitHub under its own license terms; read them before you use it at work.
Also in Training and fine-tuning
See all 41| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 86 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 85 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 84 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 83 out of 100 |