research-library-m2-v2

Metadata Category Classifier v2

Classifies paper metadata from title and abstract text into arXiv-style category labels.

Model Details

• Artifact type: full fine-tuned Transformers model
• Base model: allenai/scibert_scivocab_uncased
• Backbone type: encoder
• Model ID: M2
• Tier: T1_metadata
• Role in stack: metadata and global science-graph component
• Version: v2

This model is part of the Repository Library stack, a research system for indexing, retrieving, aligning, and reasoning over scientific papers, structured paper content, repositories, and cross-domain links between them.

Model Sources

• Hugging Face repo: https://huggingface.co/PeytonT/research-library-m2-v2
• Hugging Face collection: https://huggingface.co/collections/PeytonT/research-library-6a49c589ef4d763f7539b50d
• GitHub repository: https://github.com/peytontolbert/research_library
• Experiment config: https://github.com/peytontolbert/research_library/blob/main/models/experiments/m2_metadata_category_classifier.json
• Models directory: https://github.com/peytontolbert/research_library/tree/main/models

Intended Use

• Primary use: Classifies paper title and abstract metadata into category labels for metadata indexing and filtering.
• Downstream use: retrieval, ranking, routing, clustering, paper understanding, and global science-graph construction inside the broader Repository Library system.
• Out of scope: production safety claims, benchmark claims beyond the tracked experiment config, or deployment without task-specific validation.

Training Data

The training inputs for this package were assembled from the following Repository Library data sources:

• Source arxiv_metadata: arXiv metadata records containing titles, abstracts, authors, categories, and update metadata.

The v2 hardening pass restricts the classifier target space to a denser category set:

• max_labels: 32
• min_per_label: 16
• max_per_label: 128

Training Procedure

• Sources: arxiv_metadata
• Input fields: title, abstract
• Target fields: categories
• Filters: years [2000, 2025]
• Train/val/test split: [0.9, 0.1, 0.0]
• Max samples: 4000
• Batch size: 8
• Precision: bf16
• Objective: cross_entropy
• Learning rate: 5e-05
• Max source tokens: 512
• Max target tokens: 128
• Fine-tune strategy: full_finetune
• Number of labels: 32
• Max steps: 1000

Compute

• Hardware: 2x RTX_3090 (24 GB)
• Distributed strategy: ddp
• Estimated GPU hours in config: 0

Evaluation

Declared metrics: accuracy, macro_f1

Tracked v2 eval metrics on the current held-out split:

• eval_loss: 0.003253802889958024
• eval_accuracy: 1.0
• eval_macro_f1: 1.0
• eval_balanced_accuracy: 1.0
• eval_label_count: 31

Status: this card reflects the current tracked experiment configuration and packaged weights in the Repository Library model stack. The v2 scores are very high and should receive a leakage audit before being treated as a final benchmark.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo_id = "PeytonT/metadata-category-classifier"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

text = "TITLE: Example paper title
ABSTRACT: This paper studies representation learning for scientific documents."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
outputs = model(**inputs)
predicted_id = outputs.logits.argmax(dim=-1).item()
label = model.config.id2label.get(predicted_id, str(predicted_id))
print(label)

Limitations

• These cards are generated from tracked experiment metadata and packaged artifacts, not from a separate benchmark report or external audit.
• Several training sources are pipeline outputs from the Repository Library codebase rather than standalone public datasets.
• The v2 classifier uses a hardened 32-label target space rather than the full arXiv category universe.
• The current v2 eval result is perfect on the tracked split, which is useful as a pipeline check but should be audited for label leakage, split leakage, and class-distribution artifacts.
• These models are components of a larger research system and should be validated in their target workflow before deployment.

Project Context

• GitHub repository: https://github.com/peytontolbert/research_library
• Model collection: https://huggingface.co/collections/PeytonT/research-library-6a49c589ef4d763f7539b50d
• Publisher: PeytonT