restructuring

push docs and script

Add torchcodec install for datasets 4.x audio decoding

Prepare training stack for RunPod: env-aware notebook + bootstrap script

Align Whisper default to turbo-v3 + add document upload to Knowledge Base tab

Fix WhisperProcessor import in reload + upgrade base to large-v3-turbo

Fix mel-bin mismatch: load per-language processor from fine-tuned checkpoint

Fix jiwer crash on post-normalisation empty refs; register SLR106/105 datasets

Fix SyntaxError in Cell 17: unterminated f-string literal

Fix compute_metrics crash on empty eval references in Fula training

Fix Self-Teaching language detection: parse code from dropdown label

Add Adlam/Pular Fula integration: transliterator + 3 new datasets + normalisation pipeline

Fix language mixing: per-language prompts + Mali Bambara / Guinea Pular context

Fix model loading stuck forever + unhandled TTS crash in conversation mode

Fix: add trust_remote_code=True for datasets 3.x compatibility

Fix 'Model loading' stuck state: block in _do_asr until Whisper is ready

Fix conversation mode timeout: two-stage pipeline + faster LLM

Add real conversational memory + live learning to Conversation Mode

Fix slow build: strip runtime-irrelevant heavy packages from requirements.txt

Phase 3: Voice-to-Voice S2S pipeline — F5-TTS, LLM brain, CER metric

Fix Cell 16 ValueError: load model fp32 so AMP gradient scaler works

Fix Self-Teaching tab: float sliders, deduplication, Kaggle API fallback

Fix Kaggle trigger 401: pass credentials as env vars instead of KAGGLE_CONFIG_DIR

Fix kaggle trigger: use kaggle binary not python -m kaggle

Add kaggle to requirements.txt for Space trigger

chore: force Space restart

Fix Kaggle trigger: use kaggle CLI push instead of fake /run endpoint

Fix Wikipedia and HF harvest buttons

Fix: point Space to app.py (was app_lab.py)

Add mass training + internet self-teaching features

Align Space app with Kaggle notebook training pipeline

Fix eval_loss nan + add generation_config to Seq2SeqTrainer

Fix GenerationConfig deprecation warning in cells 13 and 16

Auto-cap MAX_STEPS based on dataset size after Cell 12 builds train_ds

Fix Cell 15: detect eval_strategy vs evaluation_strategy at runtime

Proactive notebook audit: fix 5 issues before they cause errors

Simplify Cell 2: no restart needed since PEFT is removed

Fix Cell 16 TypeError: replace PEFT LoRA with freeze/unfreeze fine-tuning

Fix Cell 16 TypeError: target encoder-only LoRA to avoid PEFT/transformers 5.x conflict

Fix Cell 14: cast input_features to fp16 in collator

Fix Cell 16: remove tokenizer arg from Seq2SeqTrainer (renamed in transformers 5.x)

Fix Cell 14: replace evaluate.load('wer') with jiwer directly

Fix Cell 13: drop bitsandbytes, load whisper-small in fp16

Fix Cell 2: restart kernel after torchvision reinstall using sentinel file

Cell 12: never raise, let Cell 12b handle no-data case gracefully

Cell 4: remove Common Voice (moved to Mozilla Data Collective, off HF)

Fix Cell 12: add diagnostics + synthetic fallback cell (12b)

Fix Cell 11: bypass WhisperProcessor entirely to avoid torchvision crash

Fix Cell 2: reinstall torchvision to match Kaggle CUDA version at runtime

Fix Cells 2+11: unpin torch, use direct WhisperProcessor import