nvidia/nemocurator-fineweb-nemotron-4-edu-classifier

NemoCurator FineWeb Nemotron-4 Edu Classifier

Model Overview

This is a text classification model designed to determine the educational value of a piece of text (score 0-5 from low to high). It is similar to the FineWeb-Edu classifier and was trained on the same text samples, but using annotations from Nemotron-4-340B-Instruct. In contrast, the original FineWeb-Edu classifier was trained using annotations from Llama 3 70B-Instruct. The NeMo Curator FineWeb Nemotron-4 Edu classifier was used as part of a classifier ensemble in the creation of the Nemotron-CC dataset. The models were finetuned starting from the Snowflake/snowflake-arctic-embed-m model.

License

GOVERNING TERMS: Use of this model is governed by the NVIDIA Open Model License Agreement. Additional Information: Apache 2.0.

References

• The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
• Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset
• Arctic-Embed: Scalable, Efficient, and Accurate Text Embedding Models

Model Architecture

• Architecture type: Transformer (BERT)
• Network architecture: Snowflake/snowflake-arctic-embed-m

How To Use in NeMo Curator

NeMo Curator improves generative AI model accuracy by processing text, image, and video data at scale for training and customization. It also provides pre-built pipelines for generating synthetic data to customize and evaluate generative AI systems.

The inference code for this model is available through the NeMo Curator GitHub repository. Check out this example notebook to get started.

How To Use in Transformers

To use the FineWeb Nemotron-4 Edu Classifier, please follow this example code:

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer


texts = ["To make lemonade, you will need lemon juice, water, and sugar."]

model = AutoModelForSequenceClassification.from_pretrained(
    "nvidia/nemocurator-fineweb-nemotron-4-edu-classifier",
    torch_dtype=torch.bfloat16,
)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

tokenizer = AutoTokenizer.from_pretrained(
    "nvidia/nemocurator-fineweb-nemotron-4-edu-classifier"
)

inputs = tokenizer(
    texts,
    return_tensors="pt",
    padding="longest",
    truncation=True,
    max_length=512,
).to(device)

with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits.squeeze(-1).float().cpu().numpy()

float_score = logits.tolist()
int_score = [int(round(max(0, min(score, 5)))) for score in logits]
pred_labels = ["high_quality" if score >= 2.5 else "low_quality" for score in logits]

print("Score:", float_score)
print("Rounded score:", int_score)
print("Predicted label:", pred_labels)
# Score: [1.0859375]
# Rounded score: [1]
# Predicted label: ['low_quality']

Input & Output

Input

• Input Type: Text
• Input Format: String
• Input Parameters: 1D
• Other Properties Related to Input: Token Limit of 512 tokens

Output

• Output Type: Classification Score
• Output Format: Float
• Output Parameters: 1D
• Other Properties Related to Output: The output range is 0-5, representing low to high educational value.

Software Integration

Runtime Engine(s):

• Python 3.10 and NeMo Curator
  

Supported Hardware Microarchitecture Compatibility:

• NVIDIA GPU, Volta™ or higher (compute capability 7.0+), CUDA 12 (or above)
  

Operating System(s):

• Ubuntu 22.04/20.04
  

Model Version(s):

• 1.0
  

Training, Testing, and Evaluation Dataset

The model was trained on the text of this dataset: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-llama3-annotations (a 467k document subset of the FineWeb dataset), with annotations coming from Nemotron-4-340B-Instruct.

Training Dataset:

Link: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-llama3-annotations

Data Collection Method by dataset

• Automated
  

Labeling Method by dataset

• Synthetic
  

Properties: The model was trained on the text of the fineweb-edu-llama3-annotations dataset, but with annotations coming from Nemotron-4-340B-Instruct instead of the provided annotations from Llama 3.1 70B. The dataset is a randomly sampled 467k document subset of the FineWeb dataset, which contains filtered documents crawled from the web. Please see https://arxiv.org/abs/2406.17557 for more details.

Evaluation Results

The models were shown to be useful in classifying high-quality content for LLM pretraining as part of an ensemble in the Nemotron-CC paper. See Table 9 from the paper below.

In the table above, "Ours-mistral" refers to the NemoCurator FineWeb Mixtral Edu Classifier, and "Ours-nemotron-340B" refers to the NemoCurator FineWeb Nemotron-4 Edu Classifier (this model). "Ours-ensembled" incudes the NemoCurator FineWeb Mixtral Edu Classifier, NemoCurator FineWeb Nemotron-4 Edu Classifier, and DCLM.

Inference

• Engine: Python 3.10 and PyTorch
• Test Hardware: NVIDIA H100

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.

Please report security vulnerabilities or NVIDIA AI Concerns here.