__pycache__/
*.pyc
.venv/
data/*
# Smoke corpus: tiny end-to-end-pipeline validator (~400 rows). Kept in
# git so the e2e SFT pipeline can be smoke-tested from a fresh clone.
!data/final-eliza1-smoke/
!data/final-eliza1-smoke/**
# Voice training data — carve out manifest + docs for voice corpora
# (e.g. same). Raw + normalized audio stays ignored (regenerable from
# upstream via scripts/voice/build_same_manifest.py). Manifests,
# READMEs, source.json, LJSpeech metadata.csv, and per-corpus .gitignore
# anchors are tracked so the corpus contract is durable in git.
!data/voice/
!data/voice/**
data/voice/**/audio/
data/voice/**/raw/
data/voice/**/ljspeech/wavs/
data/voice/**/*.wav
data/voice/**/*.flac
data/voice/**/*.mp3
data/voice/**/*.ogg
data/voice/**/*.opus
checkpoints/
# Generated benchmark outputs (parquet datasets, ablation dumps) are
# ignored, but the checked-in gate definition + engine + tests are tracked —
# the publish orchestrator reads `benchmarks/eliza1_gates.yaml` (Eliza-1
# publish-blocking eval gates) and `benchmarks/eliza1_gates.py` (the gate
# engine that turns measured eval blobs into a publish-blocking verdict).
benchmarks/*
!benchmarks/eliza1_gates.yaml
!benchmarks/eliza1_gates.py
!benchmarks/__init__.py
!benchmarks/test_eliza1_gates.py
!benchmarks/THROUGHPUT.md
!benchmarks/OPTIMIZATIONS_ROLLUP.md
!benchmarks/OPTIMIZATION_INVENTORY.md
!benchmarks/APOLLO_TUNING.md
!benchmarks/CUDA_KERNEL_PUNCHLIST.md
!benchmarks/MODELS_STATUS.md
!benchmarks/INFERENCE_OPTIMIZATION_PLAN.md
previews/
local-corpora/
wandb/
*.parquet
.vast_instance_id
.preflight.ok
.cloud_instance.json
.DS_Store
scripts/synth/scenarios/*.jsonl
# DFlash drafter distillation job outputs (see scripts/dflash/jobs/).
runs/
