Reasoning Rating Model

This repository contains the model described in the paper Meta-rater: A Multi-dimensional Data Selection Method for Pre-training Language Models.

Code: https://github.com/opendatalab/Meta-rater

Model Description

This model is a fine-tuned version of ModernBERT-base designed to evaluate the Reasoning dimension of text quality on a 5-point scale (0-5). Reasoning measures the complexity of logical thinking and analysis required from readers, focusing on multi-step reasoning, argument structure, and analytical depth.

Model Details

• Base Model: ModernBERT-base
• Parameters: 149M
• Context Window: 4,096 tokens
• Task: Text quality rating (regression)
• Score Range: 0-5 (continuous)
• Performance: 89.59% F1 score, 96.32% accuracy

Rating Scale

The model uses an additive 5-point rating system:

• 0: Contain no reasoning content
• 1: Preliminary reasoning elements with single causal relationships or simple logical judgments, lacking depth
• 2: Basic reasoning with some logical relationships requiring moderate thought, simple argumentative structures
• 3: Good reasoning complexity with multiple steps requiring complex thought, interrelated arguments with some depth
• 4: High reasoning complexity with multi-layered logic and in-depth analysis, comprehensive evaluation required
• 5: Exceptional reasoning complexity demanding deep analysis and innovative thinking, multi-dimensional and interdisciplinary

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the model and tokenizer
model_name = "opendatalab/meta-rater-reasoning-rating"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example text
text = "By analyzing the correlation between economic indicators and social outcomes, we can identify causal mechanisms that explain policy effectiveness..."

# Tokenize and predict
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=4096)
with torch.no_grad():
    outputs = model(**inputs)
    score = outputs.logits.squeeze().argmax(dim=0)

print(f"Reasoning Score: {score:.2f}")

Training Details

• Training Data: 747,422 examples from SlimPajama dataset
• Annotation Model: Llama-3.3-70B-Instruct
• Training Epochs: 10
• Evaluation Split: 93,428 test examples
• Data Split: 8:1:1 (train:dev:test)

Applications

This model is particularly valuable for:

• Educational content curation for reasoning skill development
• Research paper evaluation and academic content assessment
• Data selection for training reasoning-capable language models
• Curriculum design for critical thinking courses
• Content filtering for analytical and argumentative writing
• Quality assessment of logical and analytical texts

What the Model Evaluates

The model assesses various aspects of reasoning complexity:

• Logical structure and argument coherence
• Multi-step reasoning processes
• Causal relationship identification and analysis
• Evidence integration and synthesis
• Problem-solving approaches and methodologies
• Critical analysis depth and sophistication

Key Reasoning Indicators

• Simple reasoning: Single cause-effect relationships, basic problem-solution pairs
• Moderate reasoning: Multiple interrelated factors, comparative analysis
• Complex reasoning: Multi-layered arguments, comprehensive evaluations
• Advanced reasoning: Interdisciplinary integration, innovative thinking, mathematical models

What the Model Does NOT Consider

• The specific language the text is written in
• The length of the text
• Usage of placeholders for data privacy or safety
• Writing style, grammar, or formatting quality

Limitations

• Designed primarily for English text
• May not capture domain-specific reasoning patterns equally well across all fields
• Performance may vary for highly specialized mathematical or technical reasoning
• Should be combined with other quality metrics for comprehensive assessment

Citation

If you use this model in your research, please cite:

@article{zhuang2025meta,
  title={Meta-rater: A Multi-dimensional Data Selection Method for Pre-training Language Models},
  author={Zhuang, Xinlin and Peng, Jiahui and Ma, Ren and Wang, Yinfan and Bai, Tianyi and Wei, Xingjian and Qiu, Jiantao and Zhang, Chi and Qian, Ying and He, Conghui},
  journal={arXiv preprint arXiv:2504.14194},
  year={2025}
}

License

This model is released under the same license as the base ModernBERT model.

Contact

For questions or issues, please contact the authors or open an issue in the repository.