Com2Sense

This repo contains the dataset and baseline model weights for Com2Sense Benchmark.

It also provides access to Leaderboard submission.

---

Table of Contents

• Dataset

• Models

• Training

• Inference

• Leaderboard

---

Dataset

The directory is structured as follows:

com2sense
├── train.json
├── dev.json
└── test.json

com2sense
├── pair_id_train.json
├── pair_id_dev.son
└── pair_id_test.json

Each data file has the following format:

[   
    {
        "id": "",
        "sent": "",
        "label": "",
        "domain": "",
        "scenario": "",
        "numeracy": ""
    },
    ...
  ]

For test.json, the ground-truth labels are excluded.

Pair id files are used to get data pair information and could be used to calculate pairwise accuracy.

---

Models

Model	Std / Pair Accuracy	Weights
UnifiedQA-3B	71.31 / 51.26	Link
DeBerta-large	63.53 / 45.30	...

---

Training

For training we provide a sample script, with custom arguments (train.sh)

$ python3 main.py \
--mode train \
--dataset com2sense \
--model roberta-large \
--expt_dir ./results \
--expt_name roberta \
--run_name demo \
--seq_len 128 \
--epochs 100 \
--batch_size 16 \
--acc_step 4 \
--lr 1e-5 \
--log_interval 500 \
--gpu_ids 0,1,2,3 \
--use_amp T \
-data_parallel

The log directory for this sample script would be ./results/roberta/demo/

The Train & Validation metrics are logged to TensorBoard.

$ tensorboard --logdir ...

Note: logdir = expt_dir/expt_name/run_name/

---

Inference

TO-DO

For inference on dev set, we can modify as follows (test.sh):

$ python3 main.py \
--mode test \
--model roberta-large \
--dataset com2sense \
--ckpt ./path_to_model.pth
--test_file test \
--pred_file roberta_large_results.csv 

---

Leaderboard

To test your own model, modify the line 128:

model = Transformer(args.model, args.num_cls, text2text, num_layers=args.num_layers)

To evaluate on the official test set, we have two modes:

• Evaluation

Run with eval mode

$ python3 leaderboard.py \
--mode eval \
--ckpt /expt_dir/expt_name/run_name/model.pth

Output:

{
    'pairwise': 0.25,
    'standard': 0.50
}

• Submit

Fill in the information in submit.yaml, and then run with submit mode

$ python3 leaderboard.py \
--mode submit \
--ckpt /expt_dir/expt_name/run_name/model.pth \
--user_info ./submit.yaml

You can view the leaderboard at URL

---