Nemotron-4-340B-Instruct

$\"Model$ $\"Model$ $\"Language\"$

Model Overview

Nemotron-4-340B-Instruct is a large language model (LLM) that can be used as part of a synthetic data generation pipeline to create training data that helps researchers and developers build their own LLMs. It is a fine-tuned version of the Nemotron-4-340B-Base model, optimized for English-based single and multi-turn chat use-cases. It supports a context length of 4,096 tokens.

The base model was pre-trained on a corpus of 9 trillion tokens consisting of a diverse assortment of English based texts, 50+ natural languages, and 40+ coding languages. Subsequently the Nemotron-4-340B-Instruct model went through additional alignment steps including:

Supervised Fine-tuning (SFT)
Direct Preference Optimization (DPO)
Reward-aware Preference Optimization (RPO) (Additional in-house alignment technique)

Throughout the alignment process, we relied on only approximately 20K human-annotated data while our data generation pipeline synthesized over 98% of the data used for supervised fine-tuning and preference fine-tuning (DPO & RPO). We provide comprehensive details about our synthetic data generation pipeline in the technical report.

This results in a model that is aligned for human chat preferences, improvements in mathematical reasoning, coding and instruction-following, and is capable of generating high quality synthetic data for a variety of use cases.

Under the NVIDIA Open Model License, NVIDIA confirms:\nModels are commercially usable.\nYou are free to create and distribute Derivative Models.\nNVIDIA does not claim ownership to any outputs generated using the Models or Derivative Models.

License:

NVIDIA Open Model License

Intended use

Nemotron-4-340B-Instruct is a chat model intended for use for the English language.

Nemotron-4-340B-Instruct is designed for Synthetic Data Generation to enable developers and enterprises for building and customizing their own large language models and LLM applications.

The instruct model itself can be further customized using the NeMo Framework suite of customization tools including Parameter-Efficient Fine-Tuning (P-tuning, Adapters, LoRA, and more), and Model Alignment (SFT, SteerLM, RLHF, and more) using NeMo-Aligner.

Model Developer: NVIDIA

Model Dates: Nemotron-4-340B-Instruct was trained between December 2023 and May 2024.

Data Freshness: The pretraining data has a cutoff of June 2023.

Required Hardware

BF16 Inference:

8x H200 (1x H200 node)
16x H100 (2x H100 nodes)
16x A100 80GB (2x A100 80GB nodes)

Model Architecture:

Nemotron-4-340B-Instruct is standard decoder-only Transformer, trained with a sequence length of 4096 tokens, uses Grouped-Query Attention (GQA), and Rotary Position Embeddings (RoPE).

Architecture Type: Transformer Decoder (auto-regressive language model)

Network Architecture:\nNemotron-4

Prompt Format

Note: For Nemotron-4-340B-Instruct we recommend keeping the system prompt empty.

Single Turn

System\n\nUser\n{prompt}\nAssistant\n\">

<extra_id_0>System\n\n<extra_id_1>User\n{prompt}\n<extra_id_1>Assistant\n

Multi-Turn or Few-shot

System\n\nUser\n{prompt 1}\nAssistant\n{response 1}\nUser\n{prompt 2}\nAssistant\n{response 2}\n...\nUser\n{prompt N}\nAssistant\n\">

<extra_id_0>System\n\n<extra_id_1>User\n{prompt 1}\n<extra_id_1>Assistant\n{response 1}\n<extra_id_1>User\n{prompt 2}\n<extra_id_1>Assistant\n{response 2}\n...\n<extra_id_1>User\n{prompt N}\n<extra_id_1>Assistant\n

An example of a formattable prompt template is available in the following section.

Usage

Deployment and inference with Nemotron-4-340B-Instruct can be done in three steps using NeMo Framework:

Create a Python script to interact with the deployed model.\nCreate a Bash script to start the inference server\nSchedule a Slurm job to distribute the model across 2 nodes and associate them with the inference server.

Define the Python script call_server.py

", "", "\\x11", "User"],\n }\n sentences = text_generation(data, port=1424)['sentences']\n return sentences[0] if not batch else sentences\n\nPROMPT_TEMPLATE = """System\n\nUser\n{prompt}\nAssistant\n"""\n\nquestion = "Write a poem on NVIDIA in the style of Shakespeare"\nprompt = PROMPT_TEMPLATE.format(prompt=question)\nprint(prompt)\n\nresponse = get_generation(prompt, greedy=True, add_BOS=False, token_to_gen=1024, min_tokens=1, temp=1.0, top_p=1.0, top_k=0, repetition=1.0, batch=False)\nresponse = response[len(prompt):]\nif response.endswith(""):\n response = response[:-len("")]\nprint(response)\n\">

import json\nimport requests\n\nheaders = {\"Content-Type\": \"application/json\"}\n\ndef text_generation(data, ip='localhost', port=None):\n    resp = requests.put(f'http://{ip}:{port}/generate', data=json.dumps(data), headers=headers)\n    return resp.json()\n\n\ndef get_generation(prompt, greedy, add_BOS, token_to_gen, min_tokens, temp, top_p, top_k, repetition, batch=False):\n    data = {\n        \"sentences\": [prompt] if not batch else prompt,\n        \"tokens_to_generate\": int(token_to_gen),\n        \"temperature\": temp,\n        \"add_BOS\": add_BOS,\n        \"top_k\": top_k,\n        \"top_p\": top_p,\n        \"greedy\": greedy,\n        \"all_probs\": False,\n        \"repetition_penalty\": repetition,\n        \"min_tokens_to_generate\": int(min_tokens),\n        \"end_strings\": [\"<|endoftext|>\", \"<extra_id_1>\", \"\\x11\", \"<extra_id_1>User\"],\n    }\n    sentences = text_generation(data, port=1424)['sentences']\n    return sentences[0] if not batch else sentences\n\nPROMPT_TEMPLATE = \"\"\"<extra_id_0>System\n\n<extra_id_1>User\n{prompt}\n<extra_id_1>Assistant\n\"\"\"\n\nquestion = \"Write a poem on NVIDIA in the style of Shakespeare\"\nprompt = PROMPT_TEMPLATE.format(prompt=question)\nprint(prompt)\n\nresponse = get_generation(prompt, greedy=True, add_BOS=False, token_to_gen=1024, min_tokens=1, temp=1.0, top_p=1.0, top_k=0, repetition=1.0, batch=False)\nresponse = response[len(prompt):]\nif response.endswith(\"<extra_id_1>\"):\n    response = response[:-len(\"<extra_id_1>\")]\nprint(response)\n

Given this Python script, create a Bash script which spins up the inference server within the NeMo container (docker pull nvcr.io/nvidia/nemo:24.01.framework) and calls the Python script call_server.py. The Bash script nemo_inference.sh is as follows,

/dev/null 2>/dev/null; echo $?)\n while [ $STATUS -ne 0 ]\n do\n echo "waiting for server ($HOST:$PORT) to be up"\n sleep 10\n STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)\n done\n echo "server ($HOST:$PORT) is up running"\n}\n\n\n\n/usr/bin/python3 /opt/NeMo/examples/nlp/language_modeling/megatron_gpt_eval.py \\\n gpt_model_file=$NEMO_FILE \\\n pipeline_model_parallel_split_rank=0 \\\n server=True tensor_model_parallel_size=8 \\\n trainer.precision=bf16 pipeline_model_parallel_size=2 \\\n trainer.devices=8 \\\n trainer.num_nodes=2 \\\n web_server=False \\\n port=${WEB_PORT} &\n SERVER_PID=$!\n\n readonly local_rank="${LOCAL_RANK:=${SLURM_LOCALID:=${OMPI_COMM_WORLD_LOCAL_RANK:-}}}"\n if [ $SLURM_NODEID -eq 0 ] && [ $local_rank -eq 0 ]; then\n depends_on "0.0.0.0" ${WEB_PORT}\n\n echo "start get json"\n sleep 5\n\n echo "SLURM_NODEID: $SLURM_NODEID"\n echo "local_rank: $local_rank"\n /usr/bin/python3 /scripts/call_server.py\n echo "clean up dameons: $$"\n kill -9 $SERVER_PID\n pkill python\n fi\n wait\n\">

NEMO_FILE=$1\nWEB_PORT=1424\n\ndepends_on () {\n    HOST=$1\n    PORT=$2\n    STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)\n    while [ $STATUS -ne 0 ]\n    do\n         echo \"waiting for server ($HOST:$PORT) to be up\"\n         sleep 10\n         STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)\n    done\n    echo \"server ($HOST:$PORT) is up running\"\n}\n\n\n\n/usr/bin/python3 /opt/NeMo/examples/nlp/language_modeling/megatron_gpt_eval.py \\\n        gpt_model_file=$NEMO_FILE \\\n        pipeline_model_parallel_split_rank=0 \\\n        server=True tensor_model_parallel_size=8 \\\n        trainer.precision=bf16 pipeline_model_parallel_size=2 \\\n        trainer.devices=8 \\\n        trainer.num_nodes=2 \\\n        web_server=False \\\n        port=${WEB_PORT} &\n    SERVER_PID=$!\n\n    readonly local_rank=\"${LOCAL_RANK:=${SLURM_LOCALID:=${OMPI_COMM_WORLD_LOCAL_RANK:-}}}\"\n    if [ $SLURM_NODEID -eq 0 ] && [ $local_rank -eq 0 ]; then\n        depends_on \"0.0.0.0\" ${WEB_PORT}\n\n        echo \"start get json\"\n        sleep 5\n\n        echo \"SLURM_NODEID: $SLURM_NODEID\"\n        echo \"local_rank: $local_rank\"\n        /usr/bin/python3 /scripts/call_server.py\n        echo \"clean up dameons: $$\"\n        kill -9 $SERVER_PID\n        pkill python\n    fi\n    wait\n

3, Launch nemo_inference.sh with a Slurm script defined like below, which starts a 4-node job for model inference.

\nOUTFILE="${RESULTS}/slurm-%j-%n.out"\nERRFILE="${RESULTS}/error-%j-%n.out"\nMODEL=/Nemotron-4-340B-Instruct\n\nMOUNTS="--container-mounts=:/scripts,MODEL:/model"\nread -r -d '' cmd <

#!/bin/bash\n#SBATCH -A SLURM-ACCOUNT\n#SBATCH -p SLURM-PARITION\n#SBATCH -N 2 # number of nodes\n#SBATCH -J generation      \n#SBATCH --ntasks-per-node=8   \n#SBATCH --gpus-per-node=8\nset -x\n\nRESULTS=<PATH_TO_YOUR_SCRIPTS_FOLDER>\nOUTFILE=\"${RESULTS}/slurm-%j-%n.out\"\nERRFILE=\"${RESULTS}/error-%j-%n.out\"\nMODEL=<PATH_TO>/Nemotron-4-340B-Instruct\n\nMOUNTS=\"--container-mounts=<PATH_TO_YOUR_SCRIPTS_FOLDER>:/scripts,MODEL:/model\"\nread -r -d '' cmd <<EOF\nbash /scripts/nemo_inference.sh /model\nEOF\n\nsrun -o $OUTFILE -e $ERRFILE --container-image=\"$CONTAINER\" $MOUNTS bash -c \"${cmd}\"\n

Evaluation Results

MT-Bench (GPT-4-Turbo)

Evaluated using MT-Bench judging by GPT-4-0125-Preview as described in Appendix H in the HelpSteer2 Dataset Paper

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n