feat(stt): Add a Nemotron 3.5 ASR Finetuning Recipe

[iancarrasco-b10]

What

Adds a new training example, examples/nemotron-asr-nemo/, that fine-tunes NVIDIA's Nemotron 3.5 ASR streaming model (600M-param, multilingual, Cache-Aware FastConformer-RNNT) on Baseten.

Ported from NVIDIA's official fine-tuning recipe, following the structure/Truss conventions of the existing whisper-transformers example. Unlike Whisper, this is a NeMo recipe: it fine-tunes via NeMo's speech_to_text_finetune.py (init_from_nemo_model) using the streaming-prompt FastConformer config, driven by target_lang-tagged manifests.

Files

• config.py — Truss training job (NeMo NGC base image, 1x H100, hf_access_token secret, cache + checkpointing)
• run.sh — installs deps, clones NeMo, downloads the base .nemo checkpoint, preps data, runs the fine-tune into $BT_CHECKPOINT_DIR
• prepare_data.py — builds NeMo JSON-lines manifests (AN4 as a quick smoke test) with target_lang tags; template for bringing your own data
• requirements.txt, README.md

Notes

• Uses AN4 as an end-to-end smoke test; README documents the real path (own target_lang-tagged, punctuated manifests, tarred Lhotse shards, step budgets, replay for multilingual protection).
• Evaluation guidance uses NeMo's streaming inference at deployment latency (att_context_size).
• NEMO_BRANCH pinned to main since the streaming-prompt config is recent — pin to a release tag once available.