Model Card for harpertoken/harpertokenConvAI-finetuned

This model is a fine-tuned version of harpertoken/harpertokenConvAI, a DistilBERT-based question answering model, trained on a subset of the SQuAD dataset.

Model Details

Model Description

This is a fine-tuned question answering model based on DistilBERT, optimized for extractive QA tasks. It has been trained on a small subset of the SQuAD dataset to demonstrate fine-tuning capabilities in a CI environment.

• Developed by: bniladridas
• Model type: DistilBERT for Question Answering
• Language(s) (NLP): English
• License: MIT
• Finetuned from model: harpertoken/harpertokenConvAI

Model Sources

• Repository: https://github.com/bniladridas/harpertoken

Uses

Direct Use

This model can be used directly for question answering on passages similar to SQuAD. Provide a question and context, and it will predict the answer span.

Downstream Use

Can be further fine-tuned on domain-specific data for improved performance.

Out-of-Scope Use

Not suitable for non-English text, generative tasks, or domains outside of factual QA.

Bias, Risks, and Limitations

Trained on a limited SQuAD subset, may exhibit biases from the dataset. Performance may degrade on out-of-domain questions.

Recommendations

Evaluate on your specific data and consider additional fine-tuning for production use.

How to Get Started with the Model

from transformers import pipeline

qa = pipeline("question-answering", model="harpertoken/harpertokenConvAI-finetuned")
result = qa(question="What is the capital of France?", context="France is a country in Europe. Paris is the capital.")
print(result)

Training Details

Training Data

Subset of SQuAD 1.1 dataset (approximately 1000 examples).

Training Procedure

Training Hyperparameters

• Training regime: fp32
• Epochs: 1
• Batch size: 1
• Learning rate: 2e-5

Speeds, Sizes, Times

Trained in CI environment, minimal time due to small dataset.

Evaluation

Testing Data, Factors & Metrics

Testing Data

SQuAD validation set subset.

Metrics

F1 score, Exact Match.

Results

Basic evaluation on sample questions.

Environmental Impact

Minimal impact due to small-scale training in CI.

• Hardware Type: GitHub Actions runners
• Carbon Emitted: Negligible

Technical Specifications

Model Architecture and Objective

DistilBERT encoder with QA head for span prediction.

Compute Infrastructure

GitHub Actions Ubuntu runners.

Citation

If you use this model, please cite the original DistilBERT and SQuAD papers.

Model Card Contact

bniladridas