Skip to content

generation-numeric-tunables.md

Latest commit

 

History

History
50 lines (37 loc) · 1.69 KB

generation-numeric-tunables.md

File metadata and controls

50 lines (37 loc) · 1.69 KB
Raw

Milestone: LLM generation numeric tunables must be YAML numbers

Status: Shipped
PR: #117
Depends on: milestones/hint-segment-create-bool.md (PR #116), milestones/numeric-config-tunables.md (PR #115), milestones/generation-model-strings.md (PR #107)

Problem

narration-generate / scene-spec-generate wrap generation tunables in float() / int(). A YAML bool is a subclass of int, so:

  1. narration_from_source.temperature: true became temperature 1.0.
  2. max_context_bytes: true became a 1-byte context window.
  3. max_whisper_words_in_prompt: true became 1 (truncate the word stream).

A YAML list or string raised TypeError / ValueError at generate time, not ConfigError at load.

Goal

Fail closed at Config.from_yaml. Present values of:

  • narration_from_source.temperature / max_context_bytes
  • manim_scene_generation.temperature / max_context_bytes
  • max_whisper_segments_in_prompt / max_whisper_words_in_prompt / max_whisper_segment_text_chars

must be YAML numbers (int or float, not bool). Missing keys keep defaults.

Done when

  • Present tunables must be YAML numbers.
  • Tests for bool temperature, list max_context_bytes, quoted whisper cap, and valid numbers.
  • ruff check src/ tests/
  • pytest tests/ (677 passed, 1 skipped)
  • docgen benchmark (no clock change; meets baseline)

Out of scope

  • Nested validation numerics (ocr.sample_interval_sec, …).
  • Per-segment temperature keys that generation settings do not currently read.
  • Coercing numeric strings into numbers.