Skip to content

Latest commit

 

History

History
125 lines (100 loc) · 3.96 KB

GETTING_STARTED.md

File metadata and controls

125 lines (100 loc) · 3.96 KB

Getting started with ABI

This guide exercises the supported dependency-light alpha API. Teacher-model experiments and historical campaign reproduction require additional artifacts and should be run only from their exact protocols.

Requirements

  • Python 3.10 or newer
  • Git
  • Git LFS for a complete evidence checkout
  • Optional: a CUDA environment for extraction experiments that explicitly require it

Install from a checkout

git clone https://github.com/Yoder23/abi.git
cd abi
git lfs install
git lfs pull
python -m venv .venv
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

Activate the virtual environment using the command for your shell, or invoke its Python executable directly.

Verify the installation

python -m abi --help
python -m abi status --json
python -m abi self-check
python -m examples.segregate_capabilities

status reports the claim boundary; it is not a release certificate. self-check validates a deterministic source manifest and a domain-free English record without downloading a model. The example constructs one English record and one chemistry record and verifies their segregation.

Optional dependency groups

python -m pip install -e ".[host]"
python -m pip install -e ".[extraction]"
python -m pip install -e ".[human]"

Use host for Torch-based runtime helpers, extraction for teacher research, and human only for the frozen rating workflow. Installing a dependency group does not qualify a machine, model, or result.

Supported Python surface

The stable-for-this-alpha import surface is exported by abi/__init__.py. This minimal example creates an immutable source manifest:

from abi import build_source_model_manifest

manifest = build_source_model_manifest(
    model_id="organization/model",
    revision="a" * 40,
    revision_is_immutable=True,
    architecture="ExampleForCausalLM",
    parameter_count=1,
    tokenizer_id="organization/tokenizer",
    tokenizer_revision="a" * 40,
    license_id="example-license",
    weight_files=[
        {
            "relative_path": "model.safetensors",
            "sha256": "0" * 64,
            "bytes": 1,
        }
    ],
)
print(manifest["source_manifest_sha256"])

For a complete labeling/segregation example, read examples/segregate_capabilities.py. Phase-numbered and experiment-numbered modules are research surfaces, not stable APIs.

Run supported checks

python -m pytest -q \
  tests/test_public_release.py \
  tests/test_capability_pipeline.py \
  tests/test_capability_segregation.py
python -m ruff check abi/__init__.py abi/__main__.py \
  tests/test_public_release.py examples/segregate_capabilities.py
python scripts/check_docs.py
python -m build

The complete default research suite is intentionally stricter and requires Git LFS payloads plus exact LayerCake sibling commits. Follow the independent-review handoff. For a current development sibling, python scripts/run_review_tests.py keeps historical Phase 7 replay in an isolated exact-source fixture without modifying either checkout. See versioning and review checks.

Human review

Do not edit the sealed rating packet. Three independent humans must each complete one blinded 7,000-row form. Begin only after reading the human-rating handoff. The current count is 0/21,000; test fixtures and software validation never count as human ratings.

Common mistakes

  • Treating abi status as scientific certification.
  • Combining R7 runtime evidence with V1089 speed and R97 quality.
  • Calling a bounded ontology exhaustive capability discovery.
  • Calling exact package replay lossless foreign-teacher knowledge transfer.
  • Running an independent-hardware workflow on the development machine.
  • Ignoring missing LFS payloads or the disclosed Phase 5 tensor gap.

For claim language, use CLAIMS.md, not roadmap text.