Add scratch training notes and metadata

Use parallel hf_transfer for model uploads

Add training logs and checkpoint upload tooling

Add full dataset recovery artifacts and run notes

Add server deployment instructions and cross-platform training fixes

Add MPS training stability fixes and experiment logs

Make training scripts cross-platform (CUDA / MPS / CPU)

Add bnb_optimizer and configurable batch params to training scripts

Fix critical issues, migrate to IndicF5 fine-tuning, update pipeline

Fix Windows finetune launcher and local artifact ignores

Add auto next-batch launchers for Mac/Windows and cache-safe Mac training defaults

Add HF cache modes to safe finetune launcher and docs

Fix: remove --finetune flag (vocab size mismatch with pretrained base), train from scratch with custom Sinhala vocab

Update finetune script: epochs=20, batch_size=600, tuned for RTX 4050 6GB

Add safe finetune launcher and latest training artifacts (HF-compatible)

Add local training launch script (Mac MPS + Linux CUDA)

Add F5-TTS data prep script (Sinhala Arrow builder, bypasses pinyin)

Add start-index support to HF CC runner for non-overlapping batches

Add resilient CC batching, hybrid filtering, and 10-video training prep workflow

Upload scripts/run_cc_batches.py

Upload scripts/cc_pipeline.py

Add Gemini sentence-quality filter to CC processing pipeline

Add HF-matched CC+WAV download runner for iteration 6

Fix iteration 6 CC input parsing and UTF-8 handling

Add CC-based pipeline script (v5) — uses YouTube auto-generated captions instead of local ASR"

Batch wav2vec2 ASR inference

Fix distributed upload replacement races

Add TTS quality controls

Replace train_f5tts.py with complete fine-tuning script (vocab gen + data prep + training + inference)

Add local investigation notes and logs

Add distributed pipeline coordination

Align local pipeline with iteration 004 ASR plan

Merge origin/macos-apple-silicon into main

feat(macos): implement Apple Silicon optimizations and switch to wav2vec2 ASR

fix: switch to faster-whisper with local CT2 model for stable ASR

fix: complete ASR transcription fix in transcribe_whisper_hf

fix: resolve ASR transcription failure in whisper-hf backend

Pipeline v3: Multi-ASR backend (whisper-hf, MMS, faster-whisper), anti-hallucination, default to Lingalingeswaran/whisper-small-sinhala_v3

Pipeline v3: Multi-ASR backend (whisper-hf, MMS, faster-whisper), anti-hallucination, default to Lingalingeswaran/whisper-small-sinhala_v3

Add ASR model comparison test script (MMS, Whisper fine-tunes, Whisper large-v3)

Add scripts/train_f5tts.py

Add scripts/convert_to_f5.py

Finalize optimized pipeline with fine-tuned Whisper and model files

Finalize verified local pipeline for GPU processing

Enhance local pipeline robustness and uploader stability

Add local_pipeline.py — laptop-optimized data processing (GTX 4050 6GB VRAM)

Optimize uploader: 20-way parallel streams and 10-hour timeout

Add upload timeout and bypass stuck batch

Optimize download_and_upload.py: add batch cleanup and fix encoding

Add Phase 1: local download & upload script (Option C hybrid approach)