Major Release

Artifactiq v5.0.0 (E7): 4.6x Performance Breakthrough

Published February 3, 2026 | By Artifactiq ML Team

Executive Summary

We're excited to announce Artifactiq v5.0.0 (internally E7), our best-performing model to date. Through optimized shard-based training with 3 epochs per shard, we achieved a 4.6x improvement in mAP50 compared to our previous best model (E2/v4 baseline).

Key Achievement: mAP50 improved from 0.000316 (E2) to 0.001454 (E7) - a 4.6x improvement achieved through an important discovery: 3 epochs per shard is critical for convergence; 1 epoch causes regression.

Performance Metrics

4.6x

Improvement

0.001454

mAP50

3.5h

Training Time

Shards

Detailed Metrics

Metric	E7 (v5.0.0)	E2 (Baseline)	Improvement
mAP50	0.001454	0.000316	+360%
mAP50-95	0.000409	-	-
Precision	0.0022	-	-
Recall	0.042	-	-

The Key Discovery: Epochs Per Shard Matter

Our training experiments revealed a critical insight: the number of epochs per shard dramatically affects model convergence.

Model Comparison

Model	Shards	Epochs/Shard	mAP50	Status
E2	22	1	0.000316	Previous best
E6	46	1	0.000170	Regressed
E7	46	3	0.001454	Current best

Critical Learning: E6 (46 shards x 1 epoch) actually regressed to 0.000170 mAP50 - worse than E2. Increasing to 3 epochs per shard for E7 not only recovered but achieved 4.6x improvement over the E2 baseline.

Training Configuration

Hardware

Component	Specification
Device	Apple M4 Max
Backend	MPS (Metal Performance Shaders)
Framework	Ultralytics YOLO11m + PyTorch

Optimized Training Parameters

{
    # Shard configuration
    "shards": 46,
    "images_per_shard": 1500,
    "epochs_per_shard": 3,      # CRITICAL: Must be 3+

    # Speed optimizations
    "batch": 8,
    "mini_val": 100,            # Fast validation subset

    # Stability - Critical for MPS
    "freeze": 10,
    "cos_lr": True,
    "warmup_epochs": 3,

    # MPS Compatibility
    "workers": 0,
    "cache": False,
    "rect": False,
    "mosaic": 0.0,
    "mixup": 0.0,
    "copy_paste": 0.0,

    # Learning Rate
    "lr0": 0.001,
    "lrf": 0.01,
    "optimizer": "AdamW"
}

Key Optimizations:

freeze=10: Freezes backbone layers for stability on MPS
cos_lr=True: Cosine learning rate annealing for smoother convergence
mini_val=100: Small validation set (100 images) for fast per-shard validation
batch=8: Optimal batch size for M4 Max memory

Dataset

Item	Value
Training Images	67,991
Validation Images	7,570
Classes	39
Shard Size	1,500 images each

Shard-Based Training Architecture

Our training uses a chained shard approach where each shard's output model becomes the input for the next shard:

Base Model (federated_final.pt)
    |
    v
Shard 0 (1500 images x 3 epochs) -> shard_0/weights/last.pt
    |
    v
Shard 1 (1500 images x 3 epochs) -> shard_1/weights/last.pt
    |
    v
... (44 more shards) ...
    |
    v
Shard 45 (1500 images x 3 epochs) -> e7_best.pt

This approach allows:

Incremental training without loading entire dataset
Progress checkpoints after each shard
Real-time monitoring via ntfy.sh notifications
Resume capability if training is interrupted

Installation

Update to v5.0.0 using the Artifactiq CLI:

$ artifactiq model update
Checking for updates... v5.0.0 available (current: v4.0.0)
Downloading artifactiq-v5.0.0.onnx (76.8 MB)... done
Model updated successfully!

Or specify the version explicitly:

$ artifactiq analyze --model artifactiq:v5.0.0 --input image.jpg

Download: v5.0.0 is available on GitHub Releases

Model Files

File	Size	Format
e7_best.pt	40.5 MB	PyTorch
artifactiq-v5.0.0.onnx	76.8 MB	ONNX (deployed)

What's Next

With the E7 training breakthrough, we're planning:

E8 Training: Experiment with 5 epochs per shard to see if further improvements are possible
Cloud Training: Leverage A100 GPUs for faster iteration on hyperparameter tuning
Quantization: INT8 ONNX export for faster inference on edge devices
Expanded Classes: Adding more artifact categories based on user feedback

Changelog

Date	Event
2026-02-03	E7 released as v5.0.0
2026-02-03	E7 training complete (46 shards x 3 epochs)
2026-02-02	E6 regressed (1 epoch per shard insufficient)
2026-02-02	Optimized training: batch=8, mini_val=100, freeze=10
2026-01-31	E2 baseline established (mAP50: 0.000316)
2026-01-24	v4.0.0 released