nearlossless-context v0.2.0

nearlossless-context v0.2.0

This release makes the research prototype easier to install, run, integrate, and inspect.

Highlights

• Added a runnable end-to-end compression example:

  nearlossless-example

• Added public typed metadata contracts:

  • PrefillInfo
  • PolicyInfo
  • CompressionStats

• Documented stable metadata fields for stream, posthoc, and full modes.

• Added the py.typed marker for downstream type checkers and IDEs.

• Added offline integration tests using real Hugging Face DynamicCache objects.

• Verified compatibility with Transformers 4.57.1 through the latest supported 4.x release.

• Added clean wheel and source-distribution validation.

• Added automated PyPI publishing through GitHub Trusted Publishing.

Install

pip install nearlossless-context

Research

Preprint: https://doi.org/10.5281/zenodo.21894255

Software archive: https://doi.org/10.5281/zenodo.21894719

Full Changelog: v0.1.1...v0.2.0

nearlossless-context v0.1.1

nearlossless-context v0.1.1

First PyPI release of nearlossless-context.

Install

pip install nearlossless-context

v0.1.0 — Preprint release

nearlossless-context v0.1.0

First public research release accompanying the preprint:

Critical-Span Retention and Query-Unknown Discovery for Long-Context KV Compression

Paper DOI: https://doi.org/10.5281/zenodo.21894255

Included

• Training-free long-context KV-cache compression prototype
• Critical-span retention experiments
• Sticky surface-novelty streaming policy
• Query-hold peak/quality tradeoff
• Multi-seed retrieval and stress evaluations
• Public LongBench slice evaluation
• Reproduction scripts, results and paper source

Status

This is an experimental research prototype and alpha API, not a production inference library.

Licenses

• Software: Apache-2.0
• Paper and figures: CC BY 4.0