Sample FEVER2.0 builder docker image

The FEVER2.0 shared task requires builders to submit Docker images (via dockerhub) as part of the competition to allow for adversarial evaluation. Images must contain a single script to make predictions on a given input file using their model and host a web server (by installing the fever-api pip package) to allow for interactive evaluation as part of the breaker phase of the competition.

This repository contains an example submission based on an AllenNLP implementation of the system (see fever-allennlp). We go into depth for the following key information:

• Prediction Script
• Entrypoint
• Web Server
• Common Data

It can be run with the following commands. The first command creates a dummy container with the shared FEVER data that is used by the submission.

#Set up the data container (run once on first time)
docker create --name fever-common feverai/common

#Start a server for interactive querying of the FEVER system via the web API on port 5000
docker run --rm --volumes-from fever-common:ro -e CUDA_DEVICE=-1 -p 5000:5000 feverai/sample

#Alternatively, make predictions on a batch file and output it to `/out/predictions.jsonl` (set CUDA_DEVICE as appropriate)
docker run --rm --volumes-from fever-common:ro -e CUDA_DEVICE=-1 -v $(pwd):/out feverai/sample ./predict.sh /local/fever-common/data/fever-data/paper_dev.jsonl /out/predictions.jsonl

Shared Resources and Fair Use

The FEVER2.0 submissions will be run in a shared environment where resources will be moderated. We urge participants to ensure that these shared resources are respected.

Tensorflow users are asked to implement per-process GPU memory limits: see this post. We will set an environment variable $TF_GPU_MEMORY_FRACTION that will be tweaked for all systems in phase 2 of the shared task.

Prediction Script

The prediction script should take 2 parameters as input: the path to input file to be predicted and the path the output file to be scored:

An optional CUDA_DEVICE environment variable should be set

#!/usr/bin/env bash

default_cuda_device=0
root_dir=/local/fever-common


python -m fever.evidence.retrieve \
    --index $root_dir/data/index/fever-tfidf-ngram=2-hash=16777216-tokenizer=simple.npz \
    --database $root_dir/data/fever/fever.db \
    --in-file $1 \
    --out-file /tmp/ir.$(basename $1) \
    --max-page 5 \
    --max-sent 5

python -m allennlp.run predict \
    https://jamesthorne.co.uk/fever/fever-da.tar.gz \
    /tmp/ir.$(basename $1) \
    --output-file /tmp/labels.$(basename $1) \
    --predictor fever \
    --include-package fever.reader \
    --cuda-device ${CUDA_DEVICE:-$default_cuda_device} \
    --silent

python -m fever.submission.prepare \
    --predicted_labels /tmp/labels.$(basename $1) \
    --predicted_evidence /tmp/ir.$(basename $1) \
    --out_file $2

Entrypoint

The submission must run a flask web server to allow for interactive evaluation. In our application, the entrypoint is a function called my_sample_fever in the module sample_application (see sample_application.py). The my_sample_fever function is a factory that returns a fever_web_api object.

from fever.api.web_server import fever_web_api

def my_sample_fever(*args):
    # Set up and initialize model
    ...
    
    # A prediction function that is called by the API
    def baseline_predict(instances):
        predictions = []
        for instance in instances:
            predictions.append(...prediction for instance...)
        return predictions

    return fever_web_api(baseline_predict)

Your dockerfile can then use the waitress-serve method as the entrypoint. This will start a wsgi server calling your factory method

CMD ["waitress-serve", "--host=0.0.0.0", "--port=5000", "--call", "sample_application:my_sample_fever"]

Web Server

The web server is managed by the fever-api package. No setup or modification is required by participants. We use the default flask port of 5000 and host a single endpoint on /predict. We recommend using a client such as Postman to test your application.

POST /predict HTTP/1.1
Host: localhost:5000
Content-Type: application/json

{
	"instances":[
	    {"id":0,"claim":"this is a test claim"}, 
	    {"id":1,"claim":"this is another test claim"}, 
	]
}

API

In our sample submission, we present a simple method baseline_predict method.

   def baseline_predict(instances):
        predictions = []
        for instance in instances:
            ...prediction for instance...
            predictions.append({"predicted_label":"SUPPORTS", 
                                "predicted_evidence": [(Paris,0),(Paris,5)]})
            
        return predictions

Inputs:

• instances - a list of dictionaries containing a claim

Outputs:

• A list of dictionaries containing predicted_label (string in SUPPORTS/REFUTES/NOT ENOUGH INFO) and predicted_evidence (list of (page_name,line_number) pairs as defined in fever-scorer.

Common Data

We provide common data (the Wikipedia parse and the preprocessed data associated with the first FEVER challenge), that will be mounted in in /local/fever-common

It contains the following files (see fever.ai/resources.html for more info):

# Dataset
/local/fever-common/data/fever-data/train.jsonl
/local/fever-common/data/fever-data/paper_dev.jsonl
/local/fever-common/data/fever-data/paper_test.jsonl
/local/fever-common/data/fever-data/shared_task_dev.jsonl
/local/fever-common/data/fever-data/shared_task_test.jsonl

# Preprocessed Wikipedia Dump 
/local/fever-common/data/fever/fever.db

# Wikipedia TF-IDF Index
/local/fever-common/data/index/fever-tfidf-ngram=2-hash=16777216-tokenizer=simple.npz

# Preprocessed Wikipedia Pages (Alternative Format)
/local/fever-common/data/wiki-pages/wiki-000.jsonl
...
/local/fever-common/data/wiki-pages/wiki-109.jsonl