Getting started

Requirements

• Python 3.10–3.12 for the pinned training stack
• Git
• About 6.5 GB disk for upstream BF16 weights
• About 10–13 GB RAM for the CPU reference path; a supported GPU is preferable

Metadata, data preparation, and deterministic grading use only Python's standard library. Model loading adds PyTorch, Transformers, PEFT, Accelerate, SafeTensors, psutil, and huggingface-hub.

Install

git clone https://github.com/NT9712/tidelm.git
cd tidelm
python3 -m venv .venv
. .venv/bin/activate
pip install -e .
tidelm info

Expected current status is research with no stable adapter. That is deliberate, not a broken download.

Validate the source tree

python -m unittest discover -s tests -v
python scripts/validate-project.py
python scripts/scan-secrets.py
node scripts/build-docs.mjs --check

The online base audit is optional:

python scripts/verify-base.py

It verifies the pinned model card/config hashes, public ungated status, Apache-2.0 metadata, architecture, and 3,075,098,624 tensor parameters.

Install model runtime

pip install -e '.[train]'
tidelm download base --destination models/smollm3-3b

No model executes at package import time. The loader pins the base revision, uses SafeTensors, and sets trust_remote_code=False.

Choose a release

When a stable release exists, Releases will show its date, status, results, issues, immutable GitHub release, and download command. Until then, tidelm download adapter fails explicitly rather than substituting the base or a fake artifact.