---
language: en
license: mit
tags:
- token-classification
- ner
- math
- mathematics
- equations
base_model: roberta-base
datasets:
- lighteval/MATH-Hard
metrics:
- precision
- recall
- f1
- accuracy
pipeline_tag: token-classification
model-index:
- name: math_token_classification
  results:
  - task:
      type: token-classification
      name: Named Entity Recognition
    dataset:
      name: lighteval/MATH-Hard
      type: lighteval/MATH-Hard
      split: train
    metrics:
    - name: Accuracy
      type: accuracy
      value: 0.0
    - name: F1
      type: f1
      value: 0.0
---

# FOR TESTING PURPOSES TO SEE IF TOKEN CLASSIFICATION CAN BE USED FOR MATCHING MATH EXPRESSIONS TO MATH FIELDS

# Math Token Classification (RoBERTa-base)

This model is a fine-tuned version of **roberta-base** optimized to perform **Named Entity Recognition (NER) / Token Classification** on mathematical texts. Its primary objective is to detect inline and block LaTeX mathematical expressions within standard English text and classify them into their respective mathematical domains.

## Target Domains & Labels
The model uses a **BIO (Beginning, Inside, Outside)** tagging schema across 4 distinct mathematical disciplines:
- `ALGEBRA` (e.g., linear equations, polynomials, systems of equations)
- `CALCULUS` (e.g., integrals, derivatives, limits)
- `GEOMETRY` (e.g., shapes, coordinate planes, geometric theorems)
- `TRIGONOMETRY` (e.g., sine, cosine, identity equations)

### Label Mapping Indices
```json
{
  "0": "O",
  "1": "B-ALGEBRA", "2": "I-ALGEBRA",
  "3": "B-CALCULUS", "4": "I-CALCULUS",
  "5": "B-GEOMETRY", "6": "I-GEOMETRY",
  "7": "B-TRIGONOMETRY", "8": "I-TRIGONOMETRY"
}
```

## Intended Uses & Limitations
This model is intended for educational tools, math tutoring software, or documentation engines to help parse, analyze, or visually highlight mathematical syntax automatically.

### Limitations
- **Tokenizer Fragility:** While RoBERTa's Byte-Level BPE tokenizer handles rare math symbols effectively, highly non-standard or custom LaTeX macros not present in standard datasets might fragment predictions.
- **Context Dependence:** The expression classification depends heavily on the context words around the formula in the prompt sentence.

## Training Data & Methodology
The model was fine-tuned on **1,553 highly complex mathematical problems** extracted from the **`lighteval/MATH-Hard`** dataset (specifically Level 5 competition-grade questions). 
- **Sequence Preprocessing:** Regular expressions isolated math spans wrapped in standard delimiters (`$...$`, `$$...$$`, `\[...\]`, `\(...\)`).
- **Subword Alignment:** Character boundaries were aligned to RoBERTa token vectors via structural sequence index matching (`char_to_token`).

### Training Hyperparameters
- **Base Architecture:** RoBERTa-base (~125M parameters)
- **Learning Rate:** 3e-5
- **Batch Size:** 8
- **Epochs:** 3
- **Optimizer:** AdamW with weight decay (0.01)
- **Precision:** Mixed Precision (FP16) where available

## How to Use

You can run inference using the Hugging Face `pipeline` API:

```python
from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification

model_id = "EricAI03/math_token_classification"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

nlp_ner = pipeline("token-classification", model=model, tokenizer=tokenizer, aggregation_strategy="simple")

text = "Find the derivative of \(\int x^2\) dx and then solve the linear equation \(y = 2x + 3\)."
print(nlp_ner(text))
```