Flow: live experiment tracking for ML and GenAI | LUML
Flow<br>Live experiment tracking,<br>for ML and GenAI.<br>Local by default. Natively integrated with the LUML platform for team collaboration and the full set of platform capabilities.<br>Request demoRead the docs<br>$pip install lumlflow<br>$lumlflow ui
live<br>Runs<br>fraud_v3<br>step 100/200<br>ml<br>agent_eval<br>94% pass<br>llm<br>fraud_v2<br>0.873 auc<br>ml<br>rag_qa<br>0.78 mean score<br>llm<br>churn_xgb<br>0.741 auc<br>ml<br>agent_eval_v0<br>queued<br>llm<br>MetricsParamsModelexperiment<br>fraud_v3 · val_acc<br>0.510+0.0%
val_acctrain_loss
params · log_static7 keys<br>modellightgbm<br>learning_rate0.05<br>max_depth7<br>num_leaves63<br>early_stopping20<br>seed1337
latency<br>0ms
tokens<br>0/3.2k
cost<br>$0.012
agent_eval · spans<br>5 spans<br>spanagent.run
3.27s<br>└toolretrieve
262ms<br>└llmllm.gpt-5
2.16s<br>└tooltool.web_search
589ms<br>└spanformat.answer
458ms
scorers94% pass<br>faithfulness0.94
relevance0.88
tone0.71
Tailor this page to your workMLBothGenAI<br>Showing the unified view — ML and GenAI side by side.
One SDK<br>Classical ML and GenAI,<br>tracked the same way.<br>log_static, log_dynamic, and log_model cover training experiments. enable_tracing() plus an OTel instrumentor adds spans and eval samples for GenAI experiments. Both kinds live in the same groups.<br>ML<br>Experiments<br>· Parameters · log_static<br>· Metrics · log_dynamic<br>· Models · log_model
LLM<br>Traces & evals<br>· Spans · auto-captured<br>· Eval samples · scorers<br>· Human annotations
Charts update step by step as the script runs.<br>Spans appear in the dashboard as the agent executes.<br>One tracker, both experiment types.
flow tracker · live<br>ml + genai · one place<br>MLlightgbm · 200 steps<br>$train.py<br>log_staticlog_dynamiclog_model
GenAIrag · 312 spans<br>$agent.py
tracelog_evalannotate
tracker<br>luml·flow
unified feedlive<br>fraud_v3ml0s<br>agent_evalllm12s<br>fraud_v2ml4m<br>rag_qallm11m<br>churn_xgbml32m
5 of 142 experiments· same registry ·
two streams ─ one tracker ─ one feed
How Flow works<br>From script to dashboard in three steps.<br>Local by default.<br>Upload to LUML when an experiment is worth sharing.
01train.py<br># import once<br>from luml import tracker<br>with tracker.experiment(name="fraud_v3") as exp_id:<br>model.fit(X, y)
paramsmetricsmodeltraces
Wrap your script<br>Open a tracker.experiment(...) block around your training or eval call. Inside, log_static records parameters, log_dynamic records step metrics, and log_model captures the model. For GenAI, enable_tracing() plus an OTel instrumentor adds spans.
02$
127.0.0.1:5000
Run the local UI<br>lumlflow ui starts a dashboard at localhost:5000 that reads the local SQLite store. Compare experiments, drill into traces, and annotate eval samples without leaving your machine.
03luml/registry<br>team
fraud_v3you<br>upload<br>AB<br>churn_xgb<br>2h<br>JM<br>agent_eval<br>1d
Share with the team<br>When an experiment is worth keeping, paste your API key into the UI and click Upload. Pick organization, orbit, and collection — the model and its experiment context land in the LUML registry as a versioned artifact.
Flow UI · ML<br>ML experiments in Flow.<br>Live metrics, parameter tables, and the model artifact for every experiment.
Experiment groups<br>Group nameDescriptionTags<br>rag-pipe...tuningRAG pipeline experiments with chunk sizesllmrag
tabular...psearchXGBoost hyperparameter search on tabulartabularxgboost
image-c...icationResNet experiments on CIFAR-100visioncifar100
sentime...nalysisFine-tuning transformers on SST-2 and IMDBnlpsentiment
object-...tectionYOLO-family object detection on COCOvisiondetection
Groups — every experiment, organizedtabular-xgboost-hpsearch<br>2 Selectedmetrics.rmse 1
ExperimentDescriptionTagsDurationlogloss<br>xgb-01-d5-n300<br>california_housingxgboost0.611—<br>xgb-02-d7-n500<br>california_housingxgboost0.703—<br>xgb-05-d6-n300<br>california_housingxgboost0.8230.231<br>xgb-06-d4-n500<br>california_housingxgboost0.8960.390<br>xgb-00-d3-n100<br>california_housingxgboost0.551—<br>xgb-04-d3-n100<br>california_housingxgboost0.7730.272
Group detail — experiments side by sideOverview<br>Metrics<br>Traces<br>Evals
Parameters(11)<br>modelyolov8n<br>datasetcoco<br>image_size640<br>learning_rate0.01
Metrics(7)<br>train/box_loss0.02231<br>train/obj_loss0.02651<br>train/cls_loss0.01219
About this experiment<br>Experiment IDf6e9...0282<br>Status Completed<br>Creation2026/04/22<br>Duration0.620ms<br>Sourceseed_experiments.py<br>Tagsdetectionyolov8ncoco<br>Upload to LUML
Overview — params, metrics, modelOverview<br>Metrics<br>Traces<br>Evals
train/rmse0.497<br>step 0step 500
val/rmse0.482<br>step 0step 500
train/mae0.506<br>step 0step 500
val/mae0.497<br>step 0step 500
Metrics — step-by-step line charts
Flow UI · Unified<br>ML and GenAI experiments in Flow.<br>Same UI, with dedicated panels for each experiment type.
Overview<br>Metrics<br>Traces<br>Evals
Parameters(11)<br>modelyolov8n<br>datasetcoco<br>image_size640<br>learning_rate0.01
Metrics(7)<br>train/box_loss0.02231<br>train/obj_loss0.02651<br>train/cls_loss0.01219
About this experiment<br>Experiment IDf6e9...0282<br>Status Completed<br>Creation2026/04/22<br>Duration0.620ms<br>Sourceseed_experiments.py<br>Tagsdetectionyolov8ncoco<br>Upload to LUML
ML overview — params + metricsTrace12 spans ·...