Skip to content

docs: usability pass — gotchas page, glossary, labelled waveform, less jargon#7

Merged
shaypal5 merged 1 commit into
mainfrom
docs/usability-pass
May 12, 2026
Merged

docs: usability pass — gotchas page, glossary, labelled waveform, less jargon#7
shaypal5 merged 1 commit into
mainfrom
docs/usability-pass

Conversation

@shaypal5

Copy link
Copy Markdown
Member

Summary

Critical re-read of the docsite from the lens of a tired DS volunteer (2 hours/week, scanning for value). Reorganises content so the first scroll surfaces what a consumer actually needs — and the operator-grade detail moves out of the way without being lost.

What changed

New pages

  • docs/gotchas.md — Common mistakes. Consolidates ten previously-scattered pitfalls into one ~2-minute read: has_violence derivation rule, hardcoded speaker-dir trap, peak-is-0.79 surprise, UPPERCASE/lowercase casing, NEG-isn't-violent, split: train warning, quality_flags semantics, Google-clip flag, timestamp/pad relationship, .json vs .jsonl. Plus a clone-verification snippet.
  • docs/glossary.md — One canonical place for AGG/VIC/SW/BEN, F0, SSML, IR, ISM, dBFS, RMS, prosody cap, dirty file, weak vs strong labels, voice IDs.

New visual

  • docs/assets/sp_sv_a_0001_00_waveform.png — Real labelled waveform of an SV clip rendered from the corpus data + its .jsonl events. Now the lead element on the home page. Shows the typical escalation arc (VERBDISTPHYS) with intensity badges.

Home page

  • Leads with status pill (provisional · 2026-05-12) + a 4-line load snippet + the labelled waveform.
  • Tabbed product picker replaced with side-by-side team cards (rendered via extra.css grid).
  • Toy-corpus warning and "What's not in this corpus" callout moved below the fold.

Schema Reference

  • Rewritten as a single annotated JSON example with click-to-expand (1)! explanations.
  • Fields ordered by frequency of use: top-level → weak_labelspeakers → paths → quality_flagsacoustic_scene (Tier B) → preprocessing_appliedgeneration_metadata (collapsed) → diagnostic/reserved fields.
  • Adds .txt transcript format section (previously undocumented).

Audio Format

  • Leads with the three consumer facts (peak ~= 0.79, padding-in-timestamps, ≥3s) in one table.
  • Two-peak-fields convention surfaced (loudness_target_peak_dbfs vs normalized_dbfs).
  • Pipeline internals (M3a per-turn RMS, Stage 1/2/3 normalization, target rationale) collapsed into ??? admonitions for the curious.

Taxonomy

  • Critical admonitions opened from ??? (collapsed) to !!! (visible): NEG-trap, ACOU vs DIST.
  • New "How intensity and typology relate" section with per-typology expected max_intensity ranges.
  • Adds disclosure: scripts are LLM-generated Hebrew dialogue, not human transcripts.

Team guides

  • Reframed as the differential vs shared reference (less duplication of casing rules, schema, audio format).
  • Speaker tables split into speaker_id and voice columns (no more arrow-joined cells).
  • Full clip listings collapsed into ??? blocks; summary counts surfaced.
  • Per-team "training-time notes" trimmed to project-specific bullets.

Deliveries

  • Reordered: current state and known limitations first, diffs vs prior deliveries collapsed at the bottom.
  • Status pills for visual scan.

Operator jargon stripped

  • Milestone codes (M3a, M8a, M10a) removed from consumer-facing pages.
  • "Wet test" / "spec validation" / "pipeline bootstrapping" replaced with plain language.
  • SSML/F0/IR/ISM/Whisper/dBFS/RMS now defined on first use via Glossary cross-links.

Plumbing

  • mkdocs.yml: site_name → "AVDP Synthetic Corpus", new nav entries (Start here, Common mistakes, Glossary), abbr extension, extra_css for assets/extra.css.
  • docs/assets/extra.css: status pills + team-card grid (responsive, dark-mode aware).

Test plan

  • mkdocs build passes with zero broken anchors / missing-link warnings
  • Custom HTML (<div class=\"team-cards\">, <span class=\"status-pill\">) renders correctly through md_in_html
  • Waveform PNG resolves at /assets/sp_sv_a_0001_00_waveform.png
  • All gotchas link targets resolve (#1-dont-derive..., #4-uppercase-in-json..., etc.)
  • Auto-deploy via existing .github/workflows/docs.yml (triggers on push to main touching docs/ or mkdocs.yml)

🤖 Generated with Claude Code

…s jargon

Reorganises the docsite for a tired-volunteer reading lens. Major changes:

- New "Common mistakes" page consolidates scattered pitfalls (has_violence
  derivation, hardcoded speaker dirs, peak ~= 0.79, UPPERCASE/lowercase,
  NEG-isn't-violent, split: train, quality_flags semantics) so each one
  costs one read instead of recurring across pages.

- New Glossary page maps AGG/VIC/SW/BEN, F0, SSML, IR, ISM, dBFS, RMS,
  prosody cap, dirty file, weak vs strong labels, voice IDs.

- Home page rewritten: leads with a real labelled waveform of an SV clip
  (PNG generated from corpus data + .jsonl event boundaries), then a
  4-line load snippet, then side-by-side team cards instead of tabbed
  product picker. Toy-corpus warning and "what's not here" callout moved
  below the fold.

- Schema reference rewritten as a single annotated JSON example with
  click-to-expand explanations, ordered by frequency-of-use rather than
  by Pydantic object hierarchy. Tables retained only for EventLabel,
  manifest columns, and the .txt transcript format.

- Audio format leads with consumer facts (peak ~= 0.79, padding included
  in timestamps, two-peak-fields convention); pipeline internals (M3a
  per-turn RMS, Stage 1/2/3 normalization, target rationale) collapsed
  into optional admonitions.

- Taxonomy adds an explicit intensity-vs-typology coupling table and
  flags that scripts are LLM-generated, not human-written.

- She-Proves / Elephant pages reframed as the differential vs the
  shared reference, with cleaner speaker tables (split speaker_id /
  voice columns) and full clip listings collapsed.

- Critical ??? collapsed admonitions opened to !!! visible ones
  (peak-is-0.79, ACOU vs DIST, background event types, Tier A meaning,
  NEG-trap, casing convention). ??? reserved for skippable detail
  (per-turn RMS rationale, peak-target rationale).

- Operator jargon stripped — milestone codes (M3a/M8a/M10a), "wet
  test", "spec validation", "pipeline bootstrapping". SSML/F0/IR/ISM/
  Whisper defined on first use via Glossary cross-links.

- Custom CSS adds status pills and team cards. Logo, palette, search,
  nav structure unchanged.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@shaypal5 shaypal5 added the documentation Improvements or additions to documentation label May 12, 2026
@shaypal5 shaypal5 merged commit 9388781 into main May 12, 2026
1 of 2 checks passed
@shaypal5 shaypal5 deleted the docs/usability-pass branch May 12, 2026 21:00
@github-actions

Copy link
Copy Markdown

pr-agent-context report:

This run includes a failing check on PR #7 in repository https://github.com/DataHackIL/avdp-synth-corpus

Diagnose and fix the failing checks below, then push all of these changes in a single commit.

# Failing Checks

## FAIL-1
Type: Commit status
Context: pre-commit.ci - pr
Status: failure
URL: https://results.pre-commit.ci/run/github/1210843386/1778619602.J6z3ALh9QDCnBBTxqrEqDg

Summary:
    checks completed with failures

Run metadata:

Tool ref: v4
Tool version: 4.0.21
Trigger: pull request opened
Workflow run: 25761916157 attempt 1
Comment timestamp: 2026-05-12T21:00:28.240975+00:00
PR head commit: e27b08d1bd24d205298aa704a435a74067252616

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant