Training Pipeline

DBSprout can fine-tune a small local model on a sample of your real data, then generate from that adapter with the spec engine — without sending data to a cloud provider.

Install

pip install "dbsprout[llm]"
# plus a training backend: Unsloth (CUDA) or MLX (Apple Silicon)

End-to-End

The top-level command runs all three stages in sequence:

dbsprout train --db postgresql://localhost/myapp --sample-rows 1000 --epochs 3 --output .dbsprout

Option	Description	Default
--db	Live database URL to sample (env: DBSPROUT_TARGET_DB)	—
--sample-rows	Rows to sample (≥ 1)	1000
--epochs	Training epochs	—
--output, -o	Base directory for artifacts	.dbsprout
--seed	Sampling seed	0
--no-pii-redaction	Disable PII redaction before serialization	false
--quiet	Suppress progress output	false

Stages

Run a single stage when you need finer control:

# 1. Stratified sample from a live DB into Parquet
dbsprout train extract --db postgresql://localhost/myapp --sample-rows 1000

# 2. Serialize Parquet samples into GReaT-style JSONL
dbsprout train serialize

# 3. Fine-tune a QLoRA adapter on the serialized corpus
dbsprout train run --epochs 3

• CUDA path uses Unsloth; Apple Silicon uses MLX (auto-detected).
• Output includes a merged GGUF (Q4_K_M) adapter usable by the spec engine.

Privacy

PII values are redacted before serialization by default. Pass --no-pii-redaction only for non-sensitive data. Differential-privacy SGD (Opacus) is available on the CUDA backend; the pipeline summary reports the achieved (epsilon, delta).

Generating From the Adapter

dbsprout generate --engine spec --lora ./.dbsprout/adapter.gguf

Adapters hot-swap without restarting — see Data Generation.