ZeroPoint 1.1.0

Added

• fit and recalibrate warn (UserWarning) when fewer than 50 examples are held out, because the report's
  numbers are then too small to trust. recalibrate also warns when the temperature changes by more than a factor
  of 2 or the none level by more than 2.

Fixed

• A model made by fit kept its files in a temporary folder that was never deleted. The folder is now deleted
  when the last model that reads it is freed; save still keeps a copy wherever you want.
• The documentation said an option with depth 0 gets no probability. Depths are never exactly 0, so such an option
  gets almost no probability; the wording now says so.
• The first example in docs/usage.md and in the module docstring answered none; it is replaced by the example
  of the README, which picks an option.
• The fit time in the documentation (6 minutes) had no measurement behind it. It now gives the two measured step
  times (1.3 s and 3.4 s on 2 threads), about 4 to 12 minutes for the default 200 steps.
• The documentation shows how to install the CPU-only PyTorch before zeropoint[fit].
• Every file the training, benchmark and ablation scripts open is now closed (with blocks).
• zeropoint.fit no longer imports zeropoint.model, which removes a cyclic import; the ablation scripts drop
  unused names and a value assigned twice.
• training/zp1.py no longer claims that the model measured is the model shipped, and training/README.md
  documents a flaw in the dev split of run 1.0c: the calibration half has no written examples.

ZeroPoint 1.0.0

v1.0.0

Benchmarks: no empty excepts, files closed, one import style (CodeQL …

v0.2.0

Added

• kappa in ZeroPoint.load, and zp.kappa after loading: how fast the probability moves to none beyond
  the edge of the calibration data. The default stays at 2; 0 turns the edge off. Negative or non-numeric
  values raise ValueError.
• Decision.unseen_options: the listed options the landscape was not calibrated on, whose wells are built
  from the option's text alone.
• Decision.truncated: True when a text was longer than the 512 tokens the encoder reads.
• Documentation of κ, of long texts, and of the support landscape's weakness with overlapping queues.
• Releases carry signed SLSA level 3 provenance, and the wheel and the sdist are attached to the GitHub
  release. SECURITY.md explains how to verify them.

No decision changes: with the default settings, every text gets the same probabilities as in 0.1.0.

ZeroPoint 0.1.0

First release.

Added

• ZeroPoint.decide chooses one option from a list for a text, or returns None when none of the options
  fits. The probability of none comes from the same calculation as the probability of each option.
• ZeroPoint.decide_many for many texts at once, with one shared list of options or one list per text.
• min_confidence to return None when the chosen option is below a probability.
• Each decision holds the choice, its confidence, the probability of none, the probability and energy of
  every option, the distance beyond the edge of the calibration data, and the time it took.
• Two calibrated landscapes inside the package: everyday (150 kinds of everyday requests, calibrated on
  CLINC150; the default) and support (10 support queues, calibrated on public support tickets).
  zeropoint.landscapes() lists them.
• Two frozen encoders, multilingual-e5-small (default) and multilingual-e5-base, pinned to a fixed revision
  and run with ONNX Runtime in 8-bit or full precision. No PyTorch needed for decisions.
• ZeroPoint.fit calibrates a landscape on your own labeled texts, with no examples of none;
  save and load store it in a .safetensors file. Optional monitoring of training with ndt.
• Benchmark against similarity with the same encoder and against Laya, with the scripts to rebuild it.
• Documentation site at https://javihaus.github.io/zeropoint/.