Nemotron-Model

Nemotron SSH training and Kaggle-compatible submission workflow for the NVIDIA Nemotron reasoning challenge.

What is in this repo

• src/nemotron_model/data_bridge.py
  • prepares balanced SFT JSONL files from Kaggle train.csv
  • keeps the safer final-answer formatting we already validated locally
  • still exposes the earlier Tinker subcommands if we want that path again later
• scripts/train_trl_kaggle_sim.py
  • SSH/HPC-side LoRA or warm-start adapter training with transformers + peft + trl
  • designed for a Kaggle-like offline setup with local model paths
• scripts/build_improved_notebook.py
  • converts the original Kaggle notebook into the improved notebook version
• scripts/sync_repo_to_hpc.py
  • uploads this repo to the HFUT SSH machine over SFTP
• docs/classify_divide_trace_sft.md
  • task-by-task notes for the chosen classify-divide + stable trace + SFT approach
• slurm/train_nemotron_sft.sbatch
  • Slurm training body
• slurm/submit_train.sh
  • helper wrapper so partition/account/qos can be passed at submit time

Current remote status

The HFUT SSH endpoint is reachable and login works:

• host: 210.45.253.131
• port: 20003
• user format: u + student id

Important findings from the live probe:

• the SSH endpoint is a Slurm entry machine, not a ready-to-train GPU shell
• srun --partition=8-card --gres=gpu:1 ... currently fails for this account with:
  • invalid account or account/partition combination specified
• login node Python is only 3.6.8
• outbound container pulls from Docker Hub timed out

So this repo is ready for the SSH route, but the actual full Nemotron training still needs:

1. GPU partition access for the account
2. either a newer Python environment on the cluster, or a working container/image path

Recommended workflow

1. Prepare Kaggle data locally or on the cluster

Do not commit Kaggle competition data into Git.

python -m nemotron_model.data_bridge prepare \
  --train-csv /path/to/train.csv \
  --output-dir /path/to/prepared \
  --target-samples-per-task 1200 \
  --val-fraction 0.05 \
  --val-min-size-per-task 2

This writes:

• train_sft.jsonl
• val_sft.jsonl
• dataset_summary.json
• task_strategy_report.json
• trace_preview.md

2. Run Kaggle-style LoRA training

python scripts/train_trl_kaggle_sim.py \
  --model-path /path/to/base-model \
  --train-jsonl /path/to/prepared/train_sft.jsonl \
  --val-jsonl /path/to/prepared/val_sft.jsonl \
  --output-dir /path/to/outputs/nemotron-sft \
  --warm-start-adapter /path/to/tong-adapter \
  --max-length 4096 \
  --per-device-train-batch-size 1 \
  --gradient-accumulation-steps 16 \
  --num-train-epochs 1 \
  --learning-rate 5e-5

3. Submit through Slurm

Edit nothing inside the sbatch body unless you have to. Pass cluster-specific values at submit time:

REPO_DIR=$HOME/Nemotron-Model \
PARTITION=8-card \
QOS=duzhan \
GPU_COUNT=1 \
CPUS_PER_TASK=16 \
MEMORY=120G \
TIME_LIMIT=24:00:00 \
BASE_MODEL_PATH=/path/to/base-model \
TRAIN_JSONL=/path/to/prepared/train_sft.jsonl \
VAL_JSONL=/path/to/prepared/val_sft.jsonl \
OUTPUT_DIR=$HOME/nemotron-runs/run1 \
WARM_START_ADAPTER=/path/to/tong-adapter \
bash slurm/submit_train.sh

If your cluster admin later gives you an account name, also pass:

ACCOUNT=<your_gpu_account>

4. Sync this repo to the cluster

python scripts/sync_repo_to_hpc.py \
  --host 210.45.253.131 \
  --port 20003 \
  --user u2025171971 \
  --remote-dir /home/u2025171971/Nemotron-Model

Password can be passed with --password or HPC_PASSWORD.

Notes

• This repo intentionally excludes Kaggle raw data, adapters, and checkpoints.
• The current SSH route is prepared and tested up to remote login, file sync, and cluster probing.
• The blocking issue is cluster authorization for the GPU partition, not the training code layout.