Overview
The example catalog covers nine task families — detection, classification, language modelling, instruction tuning, speech, embeddings, medical imaging, image generation, and tool-use agents. Each one ships as a standalone job with the same FL plumbing; only the model definition and dataset wiring change.
The examples exist to show that the platform is task-agnostic. You can lift any job's layout, swap in your own model and shards, and submit through the SDK or the web wizard.
At a Glance
| Job | Model | Dataset |
|---|---|---|
| job_csgo | yolov5n-csgo (1.9 M) | CS:GO 4-class (4,262 imgs) |
| job_yolo11 | YOLO11m (20.1 M) | COCO 2017 (118k imgs) |
| job_classify | ViT-B/16 in21k (86 M) | Food-101 (101k imgs) |
| job_medical | MONAI DenseNet121 (7 M) | HAM10000 (13.4k, 7 classes) |
| job_embed | BAAI/bge-base-en-v1.5 (110 M) | all-nli (558k triplets) |
| job_speech | openai/whisper-small (244 M) | LibriSpeech train.100 (28.5k clips) |
| job_llm | microsoft/Phi-3.5-mini-instruct (3.8 B) | tatsu-lab/alpaca (52k) |
| job_diffusion | stabilityai/SDXL-base-1.0 (~3.5 B UNet) | lambdalabs/pokemon-blip-captions (833 imgs) |
| job_agent | unsloth/Llama-3.1-8B-Instruct (4-bit) | glaiveai/glaive-function-calling-v2 (113k) |
Common Job Layout
Every example job follows the same directory layout. The SDK and the web wizard expect this shape on submission:
i
All jobs share byte-identical
custom_client_executor.py and custom_persistor.py — the FL plumbing works for every task type. Only model_def.py and the framework utils change.Browse the Catalog
- job_csgo — YOLOv5 CS:GO detection. Small, fast demo job.
- job_yolo11 — YOLO11m object detection on COCO 2017.
- job_classify — ViT-B/16 image classification on Food-101.
- job_medical — MONAI DenseNet on HAM10000 dermatology.
- job_embed — BAAI/bge-base sentence embeddings on NLI triplets.
- job_speech — Whisper-small LoRA on LibriSpeech-100.
- job_llm — Phi-3.5-mini-instruct LoRA on Alpaca.
- job_diffusion — SDXL LoRA on Pokémon-BLIP captions.
- job_agent — Llama-3.1-8B-Instruct (4-bit) on Glaive function-calling.
Caveats
- Model-card metrics ≠ your FL training metrics. The published numbers on each base model's card are the original author's centralized training results. Your FL run fine-tunes from those checkpoints — final metrics depend on shard distribution, num_rounds, and local_epochs.
- Runway took down SD 1.5 (Aug 2024). Don't reference it. SDXL is the current standard for the diffusion job.
- Some base models have no published benchmarks on their HF card (e.g. Mistral-7B-Instruct-v0.3). When in doubt, pick a base that has a card you can verify.
- HAM10000 has no popular HF model card with verifiable metrics. Your FL run produces the reference.
Next Steps
- Read the Approach deep-dive for the FL plumbing every job shares.
- Submit via the SDK or the web wizard.
- See Dataset Format & Sharding for how to shape your own data.