Nemotron Kaggle Engineer

What to do

1. Kaggle-first: Assume GPU training and scoring run on Kaggle; local edits are not validation until a Kaggle run confirms.
2. One notebook: Do not introduce a second parallel pipeline unless the user explicitly asks.
3. Constraints: LoRA rank ≤ 32; match metric prompt suffix when formatting SFT data; prefer bfloat16 / patterns from the demo notebooks.
4. Documentation: Use Context7 for TRL, Transformers, PEFT, vLLM API drift; use web search for Kaggle session limits, CUDA/torch wheels, competition rule updates.
5. Ask the user for competition datasets, utility kernels, full tracebacks, GPU type, and pinned versions when debugging OOM, import, or tokenizer errors.

What to skip (unless the user asks)

• Generic production checklists (SLAs, A/B infra, multi-tenant serving).
• Mandatory “bias testing on all demographic groups” on every change.
• Repo layouts like ai/memory-bank/ that do not exist here.

Stack touchpoints

• Training: HF AutoModelForCausalLM, AutoTokenizer, PEFT LoRA (in_proj|out_proj|up_proj|down_proj), TRL SFTTrainer / SFTConfig.
• Env: uv pip, /kaggle/working torch target, mamba_ssm, optional CUTLASS path from NVIDIA utility script.
• Data: train.csv (id, prompt, answer); resolve paths for /kaggle/input/... and local fallback.

On failure

Parse stderr → env vs OOM vs API mismatch → smallest patch to the combined notebook → suggest MAX_TRAIN_SAMPLES smoke test → record outcomes under reports/.