SpeechMap data

This repository is the canonical Git history for SpeechMap model responses and production compliance analyses.

The repository was renamed from llm-compliance without rewriting history. Historical commits therefore retain the original combined code-and-data layout and object IDs. The current tree is data-focused; active collection, judging, and judge-development code lives in speechmap-eval.

Layout

• responses/: original model response rows.
• analysis/: current production compliance judgments consumed by the site.
• analysis.openai_gpt-4o-2024-11-20/: retained historical judge analyses.
• questions/: source question snapshots needed to interpret response rows.
• model_catalog.jsonl and models.txt: model/provider identity snapshots.
• backup/: retained historical response and analysis recovery files.

Judge gold sets, adjudication queues, judge-training datasets, experiment results, and judge reports belong to speechmap-eval.

The static generator and deployment source live in speechmap-site.

Local layout

The normal checkout layout is:

git/
├── speechmap-data/
├── speechmap-eval/
└── speechmap-site/

Both companion repositories default to this sibling layout. Set SPEECHMAP_DATA_ROOT to this checkout when using another location.

Data integrity

Do not treat missing or malformed rows as ordinary partial output. Response and analysis completeness, quarantine sidecars, and retained error rows must be audited with the tools in speechmap-eval before committing new model data.

Generated caches, virtual environments, remote execution state, model weights, and reproducible intermediate datasets do not belong in this repository.