sway Public

Notifications Fork 0 Star 0
fadb0ae
Branches trunk
1 Branches 0 Tags
Go to file T
Code
espadonne sway(core): NullCalibratedBackend protocol for z-score calibration
fadb0ae 47 Commits
src sway(core): NullCalibratedBackend protocol for z-score calibration
tests sway: ruff format pass
CHANGELOG.md sway: CHANGELOG for 0.1.0.dev0
LICENSE sway: scaffold standalone subproject (pyproject, LICENSE, README)
README.md sway: scaffold standalone subproject (pyproject, LICENSE, README)
pyproject.toml sway(viz): matplotlib plots for SIS, adapter ablation, KL histogram (viz extra)

dlm-sway

Differential testing for fine-tuned causal language models.

One question: did LoRA/QLoRA training actually change model behavior in a meaningful way, or is the model just defaulting to the pretrained base?

dlm-sway gives you a trustworthy, reproducible answer with eleven purpose-built primitives, each z-scored against a null-adapter baseline. No LLM judges. No external APIs. Deterministic on CPU where possible.

Install

pip install "dlm-sway[hf]"                # HuggingFace + PEFT backend
pip install "dlm-sway[hf,style,semsim]"   # full primitive battery
pip install "dlm-sway[all]"               # everything including optional viz
pip install "dlm-sway[dlm]"               # auto-generate tests from a .dlm file

90-second smoke test

dlm-sway check path/to/adapter --base HuggingFaceTB/SmolLM2-135M-Instruct

Outputs a verdict in under a minute on CPU for small models: your adapter is 4.2σ above noise ✅ or indistinguishable from a null adapter ❌.

Full suite

# sway.yaml
version: 1
models:
  base: {kind: hf, base: "HuggingFaceTB/SmolLM2-135M-Instruct"}
  ft:   {kind: hf, base: "HuggingFaceTB/SmolLM2-135M-Instruct",
         adapter: "./runs/adapter/v0003"}
suite:
  - {name: knows_concept, kind: dir,
     prompt: "The Dunning-Kruger effect describes",
     target: " a cognitive bias where",
     distractor: " a programming language"}
  - {name: no_reversion, kind: adapter_revert, paraphrases: 4}
  - {name: section_attribution, kind: section_internalization}

dlm-sway run sway.yaml              # full report to terminal + JSON
dlm-sway gate sway.yaml --junit     # CI-friendly; non-zero on fail

Why it exists

Standard benchmarks (MMLU, HellaSwag) ask "how good is this model?" That's the wrong question after a targeted LoRA fine-tune on a small user-authored document. The right question is "did the adapter actually move the model toward what I wrote?" — and existing tools answer this poorly.

dlm-sway answers it directly via eleven primitives across four categories:

Category Primitives
Adherence delta_kl, adapter_revert, prompt_collapse
Attribution section_internalization, paraphrase_invariance, preference_flip
Calibration style_fingerprint, calibration_drift, leakage
Ablation adapter_ablation ← the signature primitive

The signature primitive. adapter_ablation scales the LoRA additive term by λ ∈ {0, 0.25, 0.5, 0.75, 1.0, 1.25} and measures the divergence curve. A healthy fine-tune shows a smooth, monotonic, non-saturated response. A degenerate one shows a step function or an overshoot-then- crash. Nobody else does this because nobody else gets this close to the adapter math.

The .dlm integration

If you trained your adapter via the DocumentLanguageModel project, sway can auto-generate a test suite from your document's sections:

pip install "dlm-sway[hf,dlm]"
dlm-sway autogen path/to/doc.dlm -o sway.yaml
dlm-sway run sway.yaml

Per-section attribution tells you which parts of your document actually moved the model — a kind of signal no other tool provides.

Status

Pre-alpha. API will break. Version 0.1.0 is the first tag.

License

MIT