v0.1.4

0.1.4

• Add an English user guide with complete Python and CLI walkthroughs, result inspection, usage metrics, and batch recovery examples.
• Translate repository documentation and example messages into English.
• Update project links to TrustAIRLab/JADES_Eval and make documentation links work on both GitHub and PyPI.
• Represent multilingual detection rules and regression samples with Unicode escapes while preserving their exact runtime values.

Evaluation prompts, scoring logic, API behavior, and checkpoint compatibility are unchanged.

v0.1.3

Changelog

0.1.3

First public distribution of the JADES_light-compatible evaluator.

• Install jades-eval with pip and use the jades Python API or CLI.
• Configure each module through OpenAI-compatible endpoints, TOML and dotenv credentials.
• Track request/sample/batch time and server-reported tokens, including retries and incomplete usage.
• Resume completed samples through atomic checkpoints, exclusive output locks and a durable request ledger.
• Preserve the reference NLP, prompts, ordered scoring and weighted formula; optional fact checking is disabled by default.
• Keep textual evaluator-refusal detection warning-only and fingerprint display opt-in.
• Fix evaluator configuration drift, NLTK resource/tokenizer mismatch, search concurrency, malformed-query repair budgets and early validation of metadata and explicit env files.

Evaluator configuration is fixed at construction: create a new instance to change models, endpoints or concurrency. Recovery is per sample, not per partially completed node. See docs/p2-fixes-0.1.3.md for details.