Model training

We assume you have /workspace defined in your cluster config and are executing all commands from that folder locally. Change all commands accordingly if running on slurm or using different paths.

Download data

Get the data from HuggingFace. This might take 20-30 minutes (or more depending on your network connection) and will use ~20Gb of RAM.

import json

from datasets import load_dataset
from tqdm import tqdm

dataset = load_dataset('nvidia/OpenMathInstruct-2', split='train')

print("Converting dataset to jsonl format")
output_file = "openmathinstruct2.jsonl"
with open(output_file, 'w', encoding='utf-8') as f:
    for item in tqdm(dataset):
        f.write(json.dumps(item, ensure_ascii=False) + '\n')

print(f"Conversion complete. Output saved as {output_file}")

You can also download a subset of the data by using e.g. split='train_5M' that we used to train 70B model. See the dataset page for more details about this.

Convert to SFT format

Convert the data into the SFT format that NeMo-Aligner understands.

ns run_cmd --cluster=local \
python -m nemo_skills.training.prepare_sft_data \
    ++prompt_template=llama3-instruct \
    ++prompt_config=generic/math \
    ++preprocessed_dataset_files=/workspace/openmathinstruct2.jsonl \
    ++output_key=generated_solution \
    ++output_path=/workspace/openmathinstruct2-sft.jsonl \
    ++hf_model_name="meta-llama/Meta-Llama-3.1-8B" \
    ++filters.drop_multi_boxed=false \
    ++filters.trim_prefix=false \
    ++filters.trim_solutions=false \
    ++filters.drop_incorrect_arithmetic=false \
    ++filters.split_arithmetic=false

Prepare base model

Download the base model and convert it to NeMo format. The instructions below are for Llama3.1-8B, but the same commands should work for 70B model as well.

pip install -U "huggingface_hub[cli]"
huggingface-cli download meta-llama/Llama-3.1-8B --local-dir Llama-3.1-8B

ns convert \
    --cluster=local \
    --input_model=/workspace/Llama-3.1-8B \
    --output_model=/workspace/llama3.1-8b-nemo \
    --convert_from=hf \
    --convert_to=nemo \
    --model_type=llama \
    --num_gpus=1 \
    --hf_model_name=meta-llama/Llama-3.1-8B

Run training

Run the training (assuming slurm configuration here with the same folder structure). If your cluster has strict timeout policy, you can run multiple dependent jobs with --num_training_jobs=N.

ns train \
    --cluster=slurm \
    --expname=openmathinstruct2-repro-8b \
    --output_dir=/workspace/openmathinstruct2-repro/checkpoints \
    --nemo_model=/workspace/llama3.1-8b-nemo \
    --num_nodes=8 \
    --num_gpus=8 \
    --average_steps=10000,20000,30000,40000,50000,60000 \
    --training_data=/workspace/openmathinstruct2-sft.jsonl \
    ++model.data.train_ds.micro_batch_size=8 \
    ++model.tensor_model_parallel_size=4 \
    ++model.pipeline_model_parallel_size=1 \
    ++model.optim.lr=2e-5 \
    ++trainer.sft.save_interval=10000 \
    ++trainer.sft.max_steps=60000 \
    ++trainer.sft.max_epochs=100

For 70B model, we used 5M data subset and the following parameters, but training it longer is likely going to improve results.

ns train \
    --cluster=slurm \
    --expname=openmathinstruct2-repro-70b \
    --output_dir=/workspace/openmathinstruct2-repro-70b/checkpoints \
    --nemo_model=/workspace/llama3.1-70b-nemo \
    --num_nodes=32 \
    --num_gpus=8 \
    --average_steps=3330,6660,9990,13320,16650,20000 \
    --training_data=/workspace/openmathinstruct2-sft-5M.jsonl \
    ++model.data.train_ds.micro_batch_size=1 \
    ++model.tensor_model_parallel_size=8 \
    ++model.pipeline_model_parallel_size=2 \
    ++model.optim.lr=1e-5 \
    ++trainer.sft.save_interval=3330 \
    ++trainer.sft.max_steps=20000 \
    ++trainer.sft.max_epochs=100

If you have a job timeout, it's necessary to set the maximum time per run to 40 minutes before the timeout to allow for the final checkpoint to be saved. E.g. if your timeout is 4 hours, add ++exp_manager.max_time_per_run=00:03:20:00

If you want to follow up with checkpoint conversion and evaluation, see training docs for an example of how to do it through a convenient Python API.