TTideLM-3Bbase-stable
TideLM · 0.0.0-dev

Changelog

v0.1.0-base (2026-08-27)

Status: Stable base checkpoint — Pinned SmolLM3-3B with complete reproducible pipeline. No adapter trained; GPU required for LoRA fine-tuning.

Added

Known constraints

Unreleased

blinded comparison tooling, regression policy, CLI, tests, and security checks.

Current checkpoint status: EXP-20260826-001 failed (CPU timeout at step 3/12 after 4h). EXP-20260827-002 failed (CPU timeout during model init after 1h). EXP-20260827-003 failed (config error, seq 64 too short). EXP-20260827-004 failed (step 1/1 completed, loss=1.72, timeout post-training after 2h). EXP-20260827-005 completed (1 step, rank-1, 1.3h) but rejected — zero improvement (identical 0.6865 score). All documented in registry.

Hardware constraint: 3B BF16 training on 2-vCPU is impractical for iterative development. Even completed training yields no measurable improvement without GPU-scale compute.

No completed or promoted TideLM adapter.

The canonical machine-readable history is split across experiments/registry.json and releases/index.json. The repository-level CHANGELOG.md contains full release notes. Failed passes remain visible under Unreleased and in the experiment registry; only promoted checkpoints receive version sections.