Model Card for Model ID
This model card outlines the Pebblo Classifier, a machine learning system specialized in text classification. Developed by DAXA.AI, this model is adept at categorizing various agreement documents within organizational structures, trained on 20 distinct labels.
Model Details
Model Description
The Pebblo Classifier is a BERT-based model, fine-tuned from distilbert-base-uncased, targeting RAG (Retrieve-And-Generate) applications. It classifies text into categories such as "BOARD_MEETING_AGREEMENT," "CONSULTING_AGREEMENT," and others, streamlining document classification processes.
- Developed by: DAXA.AI
- Funded by: Open Source
- Model type: Classification model
- Language(s) (NLP): English
- License: MIT
- Finetuned from model: distilbert-base-uncased
Model Sources
- Repository: https://huggingface.co/daxa-ai/pebblo-classifier
- Demo: https://huggingface.co/spaces/daxa-ai/Daxa-Classifier
Uses
Intended Use
The model is designed for direct application in document classification, capable of immediate deployment without additional fine-tuning.
Recommendations
End-users should be cognizant of potential biases and limitations inherent in the model. For optimal use, understanding these aspects is recommended.
How to Get Started with the Model
Use the code below to get started with the model.
# Import necessary libraries
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import joblib
from huggingface_hub import hf_hub_url, cached_download
# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("daxa-ai/pebblo-classifier")
model = AutoModelForSequenceClassification.from_pretrained("daxa-ai/pebblo-classifier")
# Example text
text = "Please enter your text here."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
# Apply softmax to the logits
probabilities = torch.nn.functional.softmax(output.logits, dim=-1)
# Get the predicted label
predicted_label = torch.argmax(probabilities, dim=-1)
# URL of your Hugging Face model repository
REPO_NAME = "daxa-ai/pebblo-classifier"
# Path to the label encoder file in the repository
LABEL_ENCODER_FILE = "label_encoder.joblib"
# Construct the URL to the label encoder file
url = hf_hub_url(REPO_NAME, filename=LABEL_ENCODER_FILE)
# Download and cache the label encoder file
filename = cached_download(url)
# Load the label encoder
label_encoder = joblib.load(filename)
# Decode the predicted label
decoded_label = label_encoder.inverse_transform(predicted_label.numpy())
print(decoded_label)
Training Details
Training Data
The training dataset consists of 131,771 entries, with 20 unique labels. The labels span various document types, with instances distributed across three text sizes (128 ± x, 256 ± x, and 512 ± x words; x varies within 20). Here are the labels along with their respective counts in the dataset:
| Agreement Type | Instances |
|---|---|
| BOARD_MEETING_AGREEMENT | 4,225 |
| CONSULTING_AGREEMENT | 2,965 |
| CUSTOMER_LIST_AGREEMENT | 9,000 |
| DISTRIBUTION_PARTNER_AGREEMENT | 5,162 |
| EMPLOYEE_AGREEMENT | 3,921 |
| ENTERPRISE_AGREEMENT | 4,217 |
| ENTERPRISE_LICENSE_AGREEMENT | 9,000 |
| EXECUTIVE_SEVERANCE_AGREEMENT | 9,000 |
| FINANCIAL_REPORT_AGREEMENT | 8,381 |
| HARMFUL_ADVICE | 2,025 |
| INTERNAL_PRODUCT_ROADMAP_AGREEMENT | 7,037 |
| LOAN_AND_SECURITY_AGREEMENT | 9,000 |
| MEDICAL_ADVICE | 2,359 |
| MERGER_AGREEMENT | 7,706 |
| NDA_AGREEMENT | 5,229 |
| NORMAL_TEXT | 9,000 |
| PATENT_APPLICATION_FILLINGS_AGREEMENT | 9,000 |
| PRICE_LIST_AGREEMENT | 9,000 |
| SETTLEMENT_AGREEMENT | 3,754 |
| SEXUAL_HARRASSMENT | 8,321 |
Evaluation
Testing Data & Metrics
Testing Data
Evaluation was performed on a dataset of 82,917 entries with a temperature range of 1-1.25 for randomness. Here are the labels along with their respective counts in the dataset:
| Agreement Type | Instances |
|---|---|
| BOARD_MEETING_AGREEMENT | 4,335 |
| CONSULTING_AGREEMENT | 1,533 |
| CUSTOMER_LIST_AGREEMENT | 4,995 |
| DISTRIBUTION_PARTNER_AGREEMENT | 7,231 |
| EMPLOYEE_AGREEMENT | 1,433 |
| ENTERPRISE_AGREEMENT | 1,616 |
| ENTERPRISE_LICENSE_AGREEMENT | 8,574 |
| EXECUTIVE_SEVERANCE_AGREEMENT | 5,177 |
| FINANCIAL_REPORT_AGREEMENT | 4,264 |
| HARMFUL_ADVICE | 474 |
| INTERNAL_PRODUCT_ROADMAP_AGREEMENT | 4,116 |
| LOAN_AND_SECURITY_AGREEMENT | 6,354 |
| MEDICAL_ADVICE | 289 |
| MERGER_AGREEMENT | 7,079 |
| NDA_AGREEMENT | 1,452 |
| NORMAL_TEXT | 8,335 |
| PATENT_APPLICATION_FILLINGS_AGREEMENT | 6,177 |
| PRICE_LIST_AGREEMENT | 5,453 |
| SETTLEMENT_AGREEMENT | 5,806 |
| SEXUAL_HARRASSMENT | 4,750 |
Metrics
| Agreement Type | precision | recall | f1-score | support |
|---|---|---|---|---|
| BOARD_MEETING_AGREEMENT | 0.96 | 0.94 | 0.95 | 4335 |
| CONSULTING_AGREEMENT | 0.77 | 0.89 | 0.83 | 1533 |
| CUSTOMER_LIST_AGREEMENT | 0.84 | 0.87 | 0.85 | 4995 |
| DISTRIBUTION_PARTNER_AGREEMENT | 0.71 | 0.64 | 0.67 | 7231 |
| EMPLOYEE_AGREEMENT | 0.78 | 0.90 | 0.83 | 1433 |
| ENTERPRISE_AGREEMENT | 0.19 | 0.72 | 0.30 | 1616 |
| ENTERPRISE_LICENSE_AGREEMENT | 0.92 | 0.78 | 0.84 | 8574 |
| EXECUTIVE_SEVERANCE_AGREEMENT | 0.96 | 0.85 | 0.90 | 5177 |
| FINANCIAL_REPORT_AGREEMENT | 0.92 | 0.98 | 0.95 | 4264 |
| HARMFUL_ADVICE | 0.82 | 0.92 | 0.87 | 474 |
| INTERNAL_PRODUCT_ROADMAP_AGREEMENT | 0.94 | 0.97 | 0.96 | 4116 |
| LOAN_AND_SECURITY_AGREEMENT | 0.92 | 0.96 | 0.94 | 6354 |
| MEDICAL_ADVICE | 0.76 | 1.00 | 0.86 | 289 |
| MERGER_AGREEMENT | 0.90 | 0.55 | 0.68 | 7079 |
| NDA_AGREEMENT | 0.62 | 0.89 | 0.74 | 1452 |
| NORMAL_TEXT | 0.99 | 0.99 | 0.99 | 6049 |
| PATENT_APPLICATION_FILLINGS_AGREEMENT | 0.95 | 0.99 | 0.97 | 6177 |
| PRICE_LIST_AGREEMENT | 0.81 | 0.75 | 0.78 | 5453 |
| SETTLEMENT_AGREEMENT | 0.83 | 0.73 | 0.78 | 5806 |
| SEXUAL_HARRASSMENT | 0.98 | 0.93 | 0.96 | 4750 |
| accuracy | 0.84 | 87157 | ||
| macro avg | 0.83 | 0.86 | 0.83 | 87157 |
| weighted avg | 0.87 | 0.84 | 0.85 | 87157 |
Results
The model’s performance is summarized by precision, recall, and f1-score metrics, which are detailed across all 20 labels in the dataset. Based on the test data evaluation results, the model achieved an accuracy of 0.8376, a precision of 0.8744, and a recall of 0.8376. The F1-score, which is the harmonic mean of precision and recall, stands at 0.8478.
The evaluation loss, which measures the discrepancy between the model’s predictions and the actual values, is 0.5616. Lower loss values indicate better model performance.
The model was able to process approximately 101.886 samples per second during the evaluation, which took a total runtime of 855.4327 seconds. The model performed approximately 0.796 evaluation steps per second.
- Downloads last month
- 1,598