Universal Data Ingestion
Bring local files, Hugging Face, Kaggle, or URL sources. Normalize with adapters, map fields safely, and validate shape contracts before training.
First-Class Training Pipeline
BrewSLM pipeline: raw data to shipped SLM
- 01 Ingest HF / Kaggle / URL / local
- 02 Adapter Map schema to canonical records
- 03 Contract Check task/data compatibility
- 04 Model Select v2 introspection + memory fit
- 05 Preflight runtime/dependency guardrails
- 06 Train + Benchmark local or cloud burst
- 07 Export + Serve artifact sync + deployment
Built for Engineers
Make SLM development look like a stable software pipeline, not a notebook maze.
BrewSLM is a full stack for data contracts, capability checks, benchmark-driven model selection, multimodal runtime paths, and managed cloud burst jobs. You can stay in guided mode or drop into advanced controls at any stage.
- Reproducible steps, explicit contracts, fewer silent failures
- Works with unstructured, tabular, and instruction-style data
- One-click path for speed, deep controls for specialists
Core Product
Everything needed to ship practical SLM systems
Preflight + Capability Contracts
Catch incompatibilities early across model architecture, task type, runtime modality, dependencies, and dataset fit.
Model Selection v2
Combine curated defaults with model introspection metadata and benchmark context to make safer base-model picks.
Real Benchmark Mode
Run short benchmark jobs on sampled project data, persist results, and feed ranking decisions back into recommendations.
Multimodal Runtime Path
Text-first by default, with beta pathways for vision-language and audio-text training plus strict media preflight checks.
Cloud Burst Jobs
Managed lifecycle for remote training: submit, status, cancel, logs, and artifact sync to keep workflows consistent across local and cloud.
Product Flow
From plain intent to exportable model in five steps
01
Describe goal in plain language
BrewSLM interprets intent, suggests rewrites, and builds a safe starter plan.
02
Connect and normalize data
Adapters map diverse schemas into training-ready records with quality checks.
03
Run preflight and benchmark
Validate capability contracts and compare candidate models on your sample.
04
Launch training locally or cloud burst
Track logs, checkpoints, and metrics in one consistent control surface.
05
Export and serve
Package artifacts for deployment paths and iterate quickly with grounded feedback.
FAQ
Questions teams ask before adopting BrewSLM
Is BrewSLM only for text datasets?
No. Text is the default path, but multimodal adapters and runtime checks support vision-language and audio-text workflows.
Can beginners use this without deep ML expertise?
Yes. The Wizard maps plain intent to safe presets, guardrails block risky launches, and one-click runs handle most setup complexity.
Do I lose control if I use autopilot?
Not at all. You can inspect the generated config JSON, override any setting, and move to advanced panels at any step.
Quickstart
Get a full pipeline run locally in minutes
Use this baseline sequence to spin up backend, frontend, and worker, then launch your first training run through the Wizard or Advanced panel.
$ git clone <your-repo> __SLM__
$ cd __SLM__ && docker-compose up -d
$ cd backend && .venv/bin/celery -A app.worker worker -l info
$ cd frontend && npm run dev
Default flow emphasizes safety over raw speed.
Use `config_name` for multi-config HF datasets.
Use preflight before long jobs or cloud burst runs.
Run BrewSLM Pipeline