Stable base checkpoint — Pinned HuggingFaceTB/SmolLM3-3B@a07cc9a0... with complete reproducible pipeline.
| Property | Value |
|---|---|
| Version | v0.1.0-base |
| Date | 2026-08-27 |
| Base model | HuggingFaceTB/SmolLM3-3B |
| Revision | a07cc9a04f16550a088caea529712d1d335b0ac1 |
| Parameters | 3,075,098,624 |
| Size | 6.15 GB BF16 |
| License | Apache-2.0 |
| Status | Stable base (no adapter) |
Adapter training requires GPU — all 4 CPU experiments failed (documented in registry).
Pinned SmolLM3-3B base with complete reproducible training/evaluation pipeline. No adapter trained - requires GPU for LoRA fine-tuning. All CPU attempts failed (documented in experiment registry).
Date: 2026-08-27 · Base: HuggingFaceTB/SmolLM3-3B · Size: 3.075B parameters (6.15 GB BF16)
When a release passes the gate, the shortest verified path is:
git clone https://github.com/NT9712/tidelm.git
cd tidelm
python3 -m venv .venv && . .venv/bin/activate
pip install -e '.[train]'
tidelm download adapter --destination models/tidelm
tidelm chat --adapter models/tidelmThe downloader uses the committed release index and rejects a checksum mismatch. The upstream base downloads separately; Vercel never serves multi-gigabyte weights.
TRAIN → EVALUATE → BLIND CRITIC → REGRESSION CHECK → PROMOTE
→ VERSION → UPDATE CODE/DOCS → COMMIT → PUSH → TAG → GITHUB RELEASE
→ VERCEL DEPLOY → DEPLOYMENT/LINK/CONSOLE VERIFY → PUBLISH RESULTSA stable release includes adapter SafeTensors, PEFT config, generation config, model card, experiment report, critic/gate decision, artifact manifest, and SHA-256 checksums. Release notes list data/config changes, measured comparison, regressions, inference information, download location, and compatibility.