Skip to content

generation-zero-values.md

Latest commit

 

History

History
42 lines (31 loc) · 1.52 KB

generation-zero-values.md

File metadata and controls

42 lines (31 loc) · 1.52 KB
Raw

Milestone: honor explicit generation zeros (temperature / max_chars)

Status: Shipped
PR: #122
Depends on: milestones/generation-numeric-tunables.md (PR #117), milestones/compose-ffmpeg-timeout.md (PR #121)

Problem

Python or treats 0 / 0.0 as missing:

  1. scene-spec-generate used settings.temperature or 0.35. YAML manim_scene_generation.temperature: 0 (deterministic) became 0.35.
  2. Timing prompt enrichment used int(max_whisper_segment_text_chars or 200). YAML 0 (no truncation; same convention as the whisper count caps) became 200. Invalid types were swallowed into 200 instead of failing.

Narration-generate already passes settings.temperature through, so 0 worked there.

Goal

Honor explicit zeros. Missing / null keys keep defaults (temperature from merged scene-generation settings; max_chars 200). Present invalid max_whisper_segment_text_chars raises SceneGenerationError.

Done when

  • temperature: 0 is sent to the LLM as 0.0.
  • max_whisper_segment_text_chars: 0 does not truncate prompt text.
  • Tests cover both zeros and a bool max_chars raise.
  • ruff check src/ tests/
  • pytest tests/ (707 passed, 1 skipped)
  • docgen benchmark (no clock change; meets baseline)

Out of scope

  • Retry jitter (temperature + 0.15 * attempt after a sparse spec).
  • Changing default temperatures (merged settings still default to 0.4).