Skip to content

CHANGELOG.md

Latest commit

 

History

History
61 lines (48 loc) · 4.11 KB

CHANGELOG.md

File metadata and controls

61 lines (48 loc) · 4.11 KB
Raw

Changelog

[1.5.2] - 2026-05-22

Restructured

  • src/ split into src/standard/ + src/methodologies/ — visual separation of v1.5 production code vs v1.0 reference implementations
  • src/standard/: pipeline.py, llm_rewriter.py, translators.py (production)
  • src/methodologies/: humanizer.py, translation_chain.py, llm_rewriter.py (restored LLMRewriteProcessor), detection_pipeline.py, mixed_engine.py, postprocess.py, detectors/, utils/ (reference)
  • All import paths updated: python -m src.standard.pipeline, from src.standard import run_standard_pipeline
  • setup.py, scripts/start.*, Dockerfile, all docs, all examples updated to match

Added

  • tests/test_smoke.py — 12 smoke tests: import verification, public API surface, pure-function unit tests, showcase file integrity (all pass, no API keys required)
  • docs/README.md — documentation navigation index with audience-based routing

Fixed

  • Restored missing LLMRewriteProcessor class in src/methodologies/llm_rewriter.py (was lost during v1.5 merge — humanizer.py dispatcher referenced it but the file only had deepseek_rewrite)
  • docs/faq.md: chain description updated from DE→ES to FI (was stale from v1.5.0)
  • docs/installation.md: repo URL pointed to molly554/ai-humanize instead of lynote-ai/humanize-text
  • docs/api-reference.md: rewritten to document both Standard Pipeline and v1.0 dispatcher

[1.5.1] - 2026-05-22

Added

  • examples/showcase/ — 5 end-to-end input/output traces with all intermediate step outputs (中文改写 / 日语改写 / 一轮翻译 / 二轮翻译) plus AI-detection verdicts
  • examples/showcase/README.md — index with detection-confidence table
  • Pipeline result dict now includes output for each step (was only length)
  • New config option [pipeline].intermediate_lang to swap the intermediate translation language

Changed

  • Pipeline aligned with production reality: chain is now 4 steps (was documented as 5) — EN → ZH (DeepSeek) → JA (DeepSeek) → FI (Google) → EN (Niutrans). Finnish replaces the German→Spanish double-hop based on empirical validation.
  • README.md / README-zh.md / docs/pipeline.md updated to match the actual chain
  • Step labels now use Chinese conventions (中文改写 / 日语改写 / 一轮翻译 / 二轮翻译) for consistency with internal terminology

Why this change

The v1.5 documented chain (5 steps with German→Spanish→target) didn't match the actual production pipeline used to generate validation samples. v1.5.1 fixes this so code, docs, and showcase examples are all consistent.

[1.5.0] - 2026-05-18

Added — Standard Pipeline (Production)

  • src/pipeline.py — Production Standard Pipeline integrating Method 1 (Translation Chain) + Method 2 (LLM Rewriting) into a fixed 5-step chain
  • src/llm_rewriter.py — DeepSeek-based humanization rewriter with history-context carrying (temperature 1.3)
  • src/translators.py — Google Translate (deep-translator) + Niutrans API clients with chunking for long text
  • n8n/humanize_standard.json — Importable n8n workflow for no-code automation
  • CLI entry: python -m src.pipeline --input ... --target ...
  • Bilingual README (English + Chinese)
  • Quality metrics from expert evaluation on 50 text pairs (9.1/10 overall)

Positioning

  • v1.5 Standard Pipeline is now the recommended production path
  • v1.0 four methodologies remain in src/ as reference implementations for research and customization

[1.0.0] - 2025-12-15

Added — Four Humanization Methodologies (Reference)

  • Method 1: Translation Chain (src/translation_chain.py) — Multi-language translation hops
  • Method 2: Multi-Turn LLM Rewriting (src/humanizer.py, partial) — Iterative LLM rewrites
  • Method 3: Detection-Guided Feedback Loop (src/detection_pipeline.py, src/detectors/) — Binoculars-inspired + RoBERTa scoring loop
  • Method 4: Mixed-Engine Translation (src/mixed_engine.py) — Multiple NMT engines combined
  • Documentation of all four methodologies in docs/techniques.md
  • Comparison examples in examples/comparison/