Edit model card

Model Card for task-llm

This model supports abstractive QA tasks. Given a set of passages and a question, it tries to generate a comprehensive answer by reading the passages. In other words, the model does the generation part of retrieval augmented generation (RAG).

Model Details

This model was intended to be a T5 style multi task model trained with Bart to leverage the larger context length and better performance. At the moment, the only task supported by this model is abstractive qa

Model Description

  • Developed by: Ambika Sukla, Nlmatics Corp.
  • Model type: Generative Language Model, Abstractive QA, QASum
  • Language(s) (NLP): English
  • License: Apache 2.0
  • Finetuned from model bart:

Uses

This model supports abstractive QA tasks. Given a set of passages and a question, it tries to generate a comprehensive answer by reading the passages.

Bias, Risks, and Limitations

This model is trained with a very simple dataset and will need further fine tuning for your use cases.

Recommendations

Fine tune the model with your own data.

How to Get Started with the Model

Use the following prompt: prompt = f"###Task: abstractive_qa \n###Question: {question} \n###Passages:{passage}"

where question is your query and passage is a concatenated set of passages that needs to be considered for answering a question.

Use the code below to get started with the model:

To run this code with nlm-model-service, use the following code:

pip install nlm-utils
qa_sum_client_bart = ClassificationClient(
    model="bart",
    task="qa_sum",
    url=v100Url,
    retry=1,
)
# nlm-model-service suppports batch invocation and you can send multiple question/passage pairs at a time.
questions = ["what are the adverse reactions of Dimethylsulfoxide"]
sentences = ["Dimethylsulfoxide Adverse reactions Garlic taste in mouth, dry skin, erythema and pruritis (2), urine discoloration, halitosis, agitation, hypotension, sedation and dizziness (13) have been reported following use of DMSO. Dimethylsulfoxide Adverse reactions: malaria and loose motion."]
qa_sum_client_bart(questions, sentences)

Training Details

Training Data

Base training data was taken from this dataset with more data added for certain usage scenarios. https://github.com/microsoft/MSMARCO-Question-Answering

Training Procedure

Coming soon.

Hardware

T4, V100 or A100 GPU is recommended.

Citation

MS MARCO: A Human Generated MAchine Reading COmprehension Dataset https://arxiv.org/abs/1611.09268

BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension https://arxiv.org/abs/1910.13461

Model Card Authors

Ambika Sukla

Model Card Contact

ambika.sukla@nlmatics.com

Downloads last month
9
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.