01
Source Ingestion
Input: local files, URL, HF, Kaggle
Output: versioned raw dataset snapshot
Guardrail: malformed record and empty-split detection
Workflow
A deterministic path from source data to deployable SLM artifact.
BrewSLM turns model development into a repeatable runbook. Each stage emits explicit inputs, outputs, and checks so teams can inspect and improve the system without guessing.
Lifecycle Map
Stage-level execution model
02
Schema Normalization
Input: raw records + mapping config
Output: canonical training schema
Guardrail: field coverage and label-shape verification
03
Contract Preflight
Input: task type + model candidates + runtime profile
Output: pass/fail launch readiness report
Guardrail: capability, dependency, and memory-fit checks
04
Benchmark Sampling
Input: candidate models + sampled project data
Output: ranked model shortlist with metrics
Guardrail: score confidence floor and run reproducibility tags
05
Training Launch
Input: selected model + generated train plan
Output: checkpoints + incremental eval metrics
Guardrail: budget cap, early-stop, and retry strategy
06
Run Observability
Input: logs, metrics, worker heartbeats
Output: run timeline and issue diagnostics
Guardrail: stalled worker and metric regression alerts
07
Artifact Packaging
Input: best checkpoint + metadata + eval summary
Output: deployable artifact bundle
Guardrail: version trace and checksum validation
08
Delivery + Iteration
Input: exported bundle + serving target profile
Output: release candidate and next-run backlog
Guardrail: release gate by benchmark and drift policy
Artifact Ledger
What gets produced, where, and why it matters
| Artifact | Produced At | Used By | Purpose |
|---|---|---|---|
| Raw Snapshot | Ingest | Mapper, audits | Preserves original source state for replay |
| Canonical Dataset | Normalize | Preflight, benchmark, train | Stable schema for deterministic downstream steps |
| Preflight Report | Contract Check | Launch gate | Blocks incompatible runs before compute spend |
| Benchmark Summary | Benchmark | Model selection | Data-backed candidate ranking |
| Train Plan | Train Launch | Worker runtime | Defines training config with traceability |
| Checkpoint Set | Training | Exporter, evaluation | Captures recoverable model state |
| Export Bundle | Packaging | Serving stack | Deploy-ready model with metadata |
Operational Gates
Hard stops that prevent avoidable failure modes
Gate A: Data Integrity
Reject launches when mapping coverage, label validity, or split quality drops below threshold.
Gate B: Capability Match
Reject task/model pairs that fail compatibility contracts or runtime requirements.
Gate C: Benchmark Floor
Reject long training jobs when candidates fail minimum sampled quality criteria.
Gate D: Release Readiness
Reject export promotion when final metrics regress against previous accepted runs.
Workflow Starter
Run the full lifecycle in practical order
$ ./brewslm project create --name workflow-demo --template general
$ ./brewslm dataset import --project 1 --sample general-chat-v1
$ ./brewslm preflight --project 1 --task causal_lm --plan
$ ./brewslm train --project 1 --autopilot --one-click
$ ./brewslm export --project 1 --format huggingface --target vllm
$ ./brewslm optimize --project 1 --target mobile_iphone15