README.md

metadata

library_name: peft
license: apache-2.0
base_model: meta-llama/Llama-2-7b-chat-hf
language:
  - en
metrics:
  - f1

Overview

The model is a LoRa Adaptor based on Llama-2-7b-chat-hf. The model has been trained on a re-annotated version of the CaRB dataset.

The model produces multi-valent Open IE tuples, i.e. relations with various numbers of arguments (1, 2, or more). We provide an example below:

Consider the following sentence (taken from the CaRB dev set):

Earlier this year , President Bush made a final `` take - it - or - leave it '' offer on the minimum wage

Our model would extract the following relation from the sentence:

<President Bush, made, a final "take-it-or-leave-it" offer, on the minimum wage, earlier this year>

where we include President Bush as the subject, made as the object, a final "take-it-or-leave-it" offer as thedirect object, and on the minimum wage and earlier this year> as salient complements.

We briefly describe how to use our model in the below, and provide further details in our MulVOIEL repository on Github

Getting Started

Model Output Format

Given a sentence, the model produces textual predictions in the following format:

<subj> ,, (<auxi> ###) <predicate> ,, (<prep1> ###) <obj1>, (<prep2> ###) <obj2>, ...

How to Use

1. Install the relevant libraries as well as the MulVOIEL package:

   pip install transformers datasets peft torch
   git clone https://github.com/Teddy-Li/MulVOIEL
   cd MulVOIEL
   

2. Load the model and perform inference (example):

   from transformers import AutoModelForCausalLM, AutoTokenizer
   from peft import PeftModel
   import torch
   from llamaOIE import parse_outstr_to_triples
   from llamaOIE_dataset import prepare_input
   
   base_model_name = "meta-llama/Llama-2-7b-chat-hf"
   peft_adapter_name = "Teddy487/LLaMA2-7b-for-OpenIE"
   
   model = AutoModelForCausalLM.from_pretrained(base_model_name)
   model = PeftModel.from_pretrained(model, peft_adapter_name)
   tokenizer = AutoTokenizer.from_pretrained(base_model_name)
   
   input_text = "Earlier this year , President Bush made a final `` take - it - or - leave it '' offer on the minimum wage"
   input_text, _ = prepare_input({'s': input_text}, tokenizer, has_labels=False)
   
   input_ids = tokenizer(input_text, return_tensors="pt").input_ids
   
   outputs = model.generate(input_ids)
   outstr = tokenizer.decode(outputs[0][len(input_ids):], skip_special_tokens=True)
   triples = parse_outstr_to_triples(outstr)
   
   for tpl in triples:
       print(tpl)
   

   🍺

Model Performance

The primary benefit of our model is the ability to extract finer-grained information for predicates. On the other hand, we also report performance on a roughly comparable basis with prior SOTA open IE models, where our method is comparable and even superior to prior models, while producing finer-grained and more complex outputs. We report evaluation results in (macro) F-1 metric, as well as in the average Levenshtein Distance between gold and predicted relations:

Model	Levenshtein Distance	Macro F-1
LoRA LLaMA2-7b	5.85	50.2
LoRA LLaMA3-8b	5.04	55.3
RNN OIE *	-	49.0
IMOJIE *	-	53.5
Open IE 6 *	-	54.0/52.7

Note that the precision and recall values are not directly comparable, because we evaluate the model prediction at a finer granularity, and we use different train/dev/test arrangements as the original CaRB dataset, hence the asterisk.