Skip to content

v0.4.6 🧩

Choose a tag to compare

View all tags
@sergioburdisso sergioburdisso released this 12 Mar 13:42
· 158 commits to main since this release
v0.4.6
1515127

🚀 SDialog 0.4.6

This release note summarizes all major updates since v0.4.0 (including changes delivered across v0.4.1 → v0.4.6), presented as a single consolidated announcement.

🎙️ Audio Stack Upgrades (Major Expansion)

Since v0.4.0, the audio pipeline has been significantly extended:

  • Voice cloning support across the audio stack
  • New TTS/voice-cloning classes:
    • BaseVoiceCloneTTS
    • Qwen3TTS
    • Qwen3TTSVoiceClone
  • Persona-driven dynamic voice generation with customizable voice descriptions
  • Improved speaker-name / speaker-role mapping
  • More flexible voice assignment in the audio pipeline
  • Dialog.to_audio() improvements:
    • positional output path support
    • better optional audio path handling
  • Multiple robustness fixes (label assignment, duplicate handling, speaker case-sensitivity)
  • Updated dependencies and setup guidance for audio tooling

🧠 Agent & Tooling Enhancements

  • New @final_response_tool decorator in sdialog.agents:
    • lets an agent return tool output directly as the final answer (for non-empty outputs)
    • avoids unnecessary post-tool LLM regeneration for preformatted outputs (e.g., large tables)
  • Improved tool invocation error handling
  • Fix for orchestration behavior when agents are served
  • Progress bar improvements in dialogue generation

📊 Evaluation Improvements

Substantial upgrades to evaluation capabilities:

  • New readability and conversational metrics
  • New individual linguistic scoring classes
  • KDEDistanceEvaluator now supports multiple values per dialog
  • Improved STAR placeholder handling
  • Laplace smoothing added to dialog log-likelihood computation
  • Better divergence stability and numerical robustness:
    • fixes for division-by-zero / log(0)
    • jitter handling for zero-variance cases
  • Improved plotting flexibility (custom labels/titles)
  • ToolSequenceValidator for validating tool-usage order in dialogues

🔬 Interpretability Updates

  • Improved top-k visualization behavior and steering defaults for interpretability workflows

🧩 Core UX & API Improvements

  • Dialog.from_folder() for loading dialogs from directories
  • Prompt formatting and documentation improvements across modules
  • Expanded tutorials and usage guidance

✅ Stability & Quality

This cycle includes many targeted bug fixes across:

  • audio generation pipeline
  • evaluation math edge cases
  • agent orchestration/tool pathways
  • docs and test consistency

📚 Full Details

Full Changelog: View detailed changes

Assets 2
Loading