Build a small model you can actually audit

TideLM turns the Apache-2.0 SmolLM3-3B checkpoint into a complete model project: versioned data, real LoRA training, before/after evaluation, blinded critique, regression gates, checksummed artifacts, releases, and verified documentation.

Base stable · v0.1.0-base — Pinned SmolLM3-3B with complete pipeline. No adapter trained; GPU required for LoRA fine-tuning.

Exact, not inflated

SmolLM3 is a 3B model in the 3B–4B class, not a literal 4B model. TideLM is an independent derivative project, not a renamed upstream checkpoint.

The gauntlet

TRAIN → EVALUATE → BLIND CRITIC → REGRESSION GATE → PROMOTE
      → VERSION → PACKAGE → COMMIT → RELEASE → DEPLOY → VERIFY

A failed run remains visible and the previous stable checkpoint remains stable. No measured evidence means no improvement claim.

Start here

• Install and inspect the project
• Run inference
• Reproduce training
• Understand evaluation and gates
• Review releases
• Read limitations before use