README.md · nvidia/Nemotron-4-340B-Instruct at 3872adc4992cc9a29392ad868d358770c0926199

metadata

license: other
license_name: nvidia-open-model-license
license_link: LICENSE

Nemotron-4-340B-Instruct

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; and is a fine-tuned version of the Nemotron-4-340B-Base model, optimized for English single and multi-turn chat use-cases. 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)
Additional in-house alignment technique: Reward-aware Preference Optimization (RPO)

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:

Models are commercially usable.
You are free to create and distribute Derivative Models.
NVIDIA 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 SFT/Steer-LM/RLHF using NeMo-Aligner.

Model Developer: NVIDIA

Model Input: Text Input Format: String Input Parameters: One-Dimensional (1D)

Model Output: Text Output Format: String Output Parameters: 1D

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 (2x A100 Nodes)

FP8 Inference:

8x H100 (1x H100 Node)

Model Architecture:

Nemotron-4-340B-Base, 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: Nemotron-4

Usage

We will spin up an inference server and then call the inference server in a python script. Let’s first define the python script call_server.py

import json
import requests

headers = {"Content-Type": "application/json"}

def text_generation(data, ip='localhost', port=None):
    resp = requests.put(f'http://{ip}:{port}/generate', data=json.dumps(data), headers=headers)
    return resp.json()


def get_generation(prompt, greedy, add_BOS, token_to_gen, min_tokens, temp, top_p, top_k, repetition, batch=False):
    data = {
        "sentences": [prompt] if not batch else prompt,
        "tokens_to_generate": int(token_to_gen),
        "temperature": temp,
        "add_BOS": add_BOS,
        "top_k": top_k,
        "top_p": top_p,
        "greedy": greedy,
        "all_probs": False,
        "repetition_penalty": repetition,
        "min_tokens_to_generate": int(min_tokens),
        "end_strings": ["<|endoftext|>", "<extra_id_1>", "\x11", "<extra_id_1>User"],
    }
    sentences = text_generation(data, port=1424)['sentences']
    return sentences[0] if not batch else sentences

PROMPT_TEMPLATE = """<extra_id_0>System

<extra_id_1>User
{prompt}
<extra_id_1>Assistant
"""

question = "Write a poem on NVIDIA in the style of Shakespeare"
prompt = PROMPT_TEMPLATE.format(prompt=question)
print(prompt)

response = 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)
response = response[len(prompt):]
if response.endswith("<extra_id_1>"):
    response = response[:-len("<extra_id_1>")]
print(response)

Given this python script, we will 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,

NEMO_FILE=$1
WEB_PORT=1424

depends_on () {
    HOST=$1
    PORT=$2
    STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)
    while [ $STATUS -ne 0 ]
    do
         echo "waiting for server ($HOST:$PORT) to be up"
         sleep 10
         STATUS=$(curl -X PUT http://$HOST:$PORT >/dev/null 2>/dev/null; echo $?)
    done
    echo "server ($HOST:$PORT) is up running"
}



/usr/bin/python3 /opt/NeMo/examples/nlp/language_modeling/megatron_gpt_eval.py \
        gpt_model_file=$NEMO_FILE \
        pipeline_model_parallel_split_rank=0 \
        server=True tensor_model_parallel_size=8 \
        trainer.precision=bf16 pipeline_model_parallel_size=2 \
        trainer.devices=8 \
        trainer.num_nodes=2 \
        web_server=False \
        port=${WEB_PORT} &
    SERVER_PID=$!

    readonly local_rank="${LOCAL_RANK:=${SLURM_LOCALID:=${OMPI_COMM_WORLD_LOCAL_RANK:-}}}"
    if [ $SLURM_NODEID -eq 0 ] && [ $local_rank -eq 0 ]; then
        depends_on "0.0.0.0" ${WEB_PORT}

        echo "start get json"
        sleep 5

        echo "SLURM_NODEID: $SLURM_NODEID"
        echo "local_rank: $local_rank"
        /usr/bin/python3 /scripts/call_server.py
        echo "clean up dameons: $$"
        kill -9 $SERVER_PID
        pkill python
    fi
    wait

3, We can launch the nemo_inferece.sh with a slurm script defined like below, which starts a 4-node job for the model inference.

#!/bin/bash
#SBATCH -A SLURM-ACCOUNT
#SBATCH -p SLURM-PARITION
#SBATCH -N 2 # number of nodes
#SBATCH -J generation      
#SBATCH --ntasks-per-node=8   
#SBATCH --gpus-per-node=8
set -x

RESULTS=<PATH_TO_YOUR_SCRIPTS_FOLDER>
OUTFILE="${RESULTS}/slurm-%j-%n.out"
ERRFILE="${RESULTS}/error-%j-%n.out"
MODEL=<PATH_TO>/Nemotron-4-340B-Instruct

MOUNTS="--container-mounts=<PATH_TO_YOUR_SCRIPTS_FOLDER>:/scripts,MODEL:/model"
read -r -d '' cmd <<EOF
bash /scripts/nemo_inference.sh /model
EOF

srun -o $OUTFILE -e $ERRFILE --container-image="$CONTAINER" $MOUNTS bash -c "${cmd}"

Evaluation Results

MT-Bench (GPT-4-Turbo)

Evaluated using MT-Bench judging by GPT-4-Turbo as described in the HelpSteer2 Dataset Paper.

Please note that these numbers aren't comparable with original MT-bench as we corrected some references

total	writing	roleplay	extraction	stem	humanities	reasoning	math	coding	turn 1	turn 2
8.22	8.70	8.70	9.20	8.75	8.95	6.40	8.40	6.70	8.61	7.84

IFEval

Evaluated using the Instruction Following Eval (IFEval) introduced in Instruction-Following Evaluation for Large Language Models.

Prompt-Strict Acc	Instruction-Strict Acc
79.9	86.1

MMLU

Evaluated using the Multi-task Language Understanding benchmarks as introduced in Measuring Massive Multitask Language Understanding

MMLU 0-shot
78.7

GSM8K

Evaluated using the Grade School Math 8K (GSM8K) benchmark as introduced in Training Verifiers to Solve Math Word Problems.

GSM8K 0-shot
92.3

HumanEval

Evaluated using the HumanEval benchmark as introduced in Evaluating Large Language Models Trained on Code.

HumanEval 0-shot
73.2

MBPP

Evaluated using the MBPP Dataset as introduced in the Program Synthesis with Large Language Models paper.

MBPP 0-shot
75.4

Arena Hard

Evaluated using the Arena-Hard Pipeline from the LMSys Org.

Arena Hard
54.2

AlpacaEval 2.0 LC

Evaluated using the AlpacaEval 2.0 LC (Length Controlled) as introduced in the paper: Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators

AlpacaEval 2.0 LC
41.5

TFEval

Evaluated using the CantTalkAboutThis Dataset as introduced in the CantTalkAboutThis: Aligning Language Models to Stay on Topic in Dialogues paper.

Distractor F1	On-topic F1
81.7	97.7

Adversarial Testing and Red Teaming Efforts

The Nemotron-4 340B-Instruct model underwent extensive safety evaluation including adversarial testing via three distinct methods:

Garak, is an automated LLM vulnerability scanner that probes for common weaknesses, including prompt injection and data leakage.
AEGIS, is a content safety evaluation dataset and LLM based content safety classifier model, that adheres to a broad taxonomy of 13 categories of critical risks in human-LLM interactions.
Human Content Red Teaming leveraging human interaction and evaluation of the models' responses.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards [Insert Link to Model Card++ here]. Please report security vulnerabilities or NVIDIA AI Concerns here.