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:
1qa_sum_client_bart = ClassificationClient(
2 model="bart",
3 task="qa_sum",
4 url=v100Url,
5 retry=1,
6)
7# nlm-model-service suppports batch invocation and you can send multiple question/passage pairs at a time.
8questions = ["what are the adverse reactions of Dimethylsulfoxide"]
9sentences = ["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."]
10qa_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-AnsweringTraining 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/1611.09268BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
https://arxiv.org/abs/1910.13461Model Card Authors
Ambika Sukla
Model Card Contact
