sway Public

Watch 0 Fork 0 Star 0
markdown · 2237 bytes Raw Blame History

Changelog

0.1.0.dev0 — 2026-04-20

Initial pre-alpha. Full 11-primitive battery shipped.

Primitives

  • Adherence
    • delta_kl — mean JS/KL divergence between base and fine-tuned next-token distributions
    • adapter_revert — reversion under adversarial paraphrase (needs sway-eval[semsim])
    • prompt_collapse — exponential-decay fit of divergence over context length
  • Attribution
    • section_internalization (flagship) — per-section effective_sis with leak check
    • paraphrase_invariance — memorization vs. generalization, intent-aware
    • preference_flip — DPO/ORPO chosen/rejected margin inversion
  • Calibration
    • style_fingerprint — 6-dim numpy-only stylistic shift vs. document
    • calibration_drift — general-knowledge regression on a packaged 30-item pack
    • leakage — greedy LCS recall + perturbation fragility
  • Ablation
    • adapter_ablation (signature primitive) — λ-scaled divergence curve with linearity, saturation, overshoot metrics
  • Baseline
    • null_adapter — stats scaffolding for z-score calibration (implementation pending)

Infrastructure

  • DifferentialBackend + ScalableDifferentialBackend protocols
  • HuggingFace + PEFT backend with disable_adapter / set_adapter toggling and LoRA-scale mutation
  • Dummy backend for unit tests (canned responses + linear-blend scalable mode)
  • YAML spec loader, composite score (four-category weighted), rich terminal + JSON + JUnit + Markdown reports
  • Typer CLI: run, gate, check, diff, autogen, doctor, report
  • .dlm bridge (dlm-sway[dlm]): resolver + full-battery autogen
  • Matplotlib visualizations (dlm-sway[viz]): SIS bar chart, ablation curve, KL histogram

Known gaps

  • Null-adapter baseline is scaffolded but its HF-level materialization (building random-init LoRAs at matched rank) is not yet wired — probes fall back to fixed thresholds until the next milestone.
  • Custom backend entry-point dispatch (kind: custom) is stubbed but not implemented.
  • MLX backend is registered as a future-milestone target; all MLX paths raise BackendNotAvailableError.
  • PyPI publication of the dlm-sway wheel is pending a clean CI release workflow.
View source
1 # Changelog
2
3 ## 0.1.0.dev0 — 2026-04-20
4
5 Initial pre-alpha. Full 11-primitive battery shipped.
6
7 ### Primitives
8
9 - **Adherence**
10 - `delta_kl` — mean JS/KL divergence between base and fine-tuned next-token distributions
11 - `adapter_revert` — reversion under adversarial paraphrase (needs `sway-eval[semsim]`)
12 - `prompt_collapse` — exponential-decay fit of divergence over context length
13 - **Attribution**
14 - `section_internalization` *(flagship)* — per-section `effective_sis` with leak check
15 - `paraphrase_invariance` — memorization vs. generalization, intent-aware
16 - `preference_flip` — DPO/ORPO chosen/rejected margin inversion
17 - **Calibration**
18 - `style_fingerprint` — 6-dim numpy-only stylistic shift vs. document
19 - `calibration_drift` — general-knowledge regression on a packaged 30-item pack
20 - `leakage` — greedy LCS recall + perturbation fragility
21 - **Ablation**
22 - `adapter_ablation` *(signature primitive)* — λ-scaled divergence curve with linearity, saturation, overshoot metrics
23 - **Baseline**
24 - `null_adapter` — stats scaffolding for z-score calibration (implementation pending)
25
26 ### Infrastructure
27
28 - `DifferentialBackend` + `ScalableDifferentialBackend` protocols
29 - HuggingFace + PEFT backend with `disable_adapter` / `set_adapter` toggling and LoRA-scale mutation
30 - Dummy backend for unit tests (canned responses + linear-blend scalable mode)
31 - YAML spec loader, composite score (four-category weighted), rich terminal + JSON + JUnit + Markdown reports
32 - Typer CLI: `run`, `gate`, `check`, `diff`, `autogen`, `doctor`, `report`
33 - `.dlm` bridge (`dlm-sway[dlm]`): resolver + full-battery autogen
34 - Matplotlib visualizations (`dlm-sway[viz]`): SIS bar chart, ablation curve, KL histogram
35
36 ### Known gaps
37
38 - Null-adapter baseline is scaffolded but its HF-level materialization (building random-init LoRAs at matched rank) is not yet wired — probes fall back to fixed thresholds until the next milestone.
39 - Custom backend entry-point dispatch (`kind: custom`) is stubbed but not implemented.
40 - MLX backend is registered as a future-milestone target; all MLX paths raise `BackendNotAvailableError`.
41 - PyPI publication of the `dlm-sway` wheel is pending a clean CI release workflow.