Awesome

Tk-Instruct

• This repo releases our implementation for the Tk-Instruct model in the Super-NaturalInstructions paper.
• Tk-Instruct is a preliminary attempt towards general-purpose AI that can solve many NLP tasks by following in-context instructions (plain language task definitions or k-shot examples).
• It is built based on the pretrained T5 model, and finetuned on our data.
• You can play with the 11B model via our online demo!

Requirements

Our main experiments and analysis are conducted on the following environment:

• CUDA (11.3)
• cuDNN (8.2.0.53)
• Pytorch (1.10.0)
• Transformers (4.17.0)
• DeepSpeed

You can refer to the Dockerfile for setting up the environment and install the required python libraries by running

pip install -r requirements.txt

Note: after the main exploration with 3B model, we train our 11B model on TPUs using the T5 code here.

Data

Our models are trained and evaluated on Super-NaturalInstructions, which can be cloned by running:

git clone git@github.com:allenai/natural-instructions.git data

Since Super-NaturalInstructions didn't provide an official split for the development set, in order to do evaluation during training time, you can mannualy create a dev_tasks.txt in the data/splits/default folder. We found it unclear what should be a meaningful validation set, under such cross-task generalization setting. You can use a part of the training tasks for validation, or you can set apart tasks in some categories for validation.

If you want to use the T5 code here, you can convert the data into text2text format with scripts/convert_data_to_s2s.sh.

Training

A sample script for training the Tk-Instruct 3B model in our paper can be found at scripts/train_tk_instruct.sh. You can run it as follows:

./scripts/train_tk_instruct.sh

However, if you are familiar with Beaker, you can refer to the beaker_configs/default_experiment.yaml for a sample experiment config, and modifying src/create_exps.py to easily starts a set of experiments by running:

python src/create_exps.py

Released Checkpoints

Our 3B and 11B model checkpoints are accessible via the Hugging Face Hub. You can load them easily using the Transformers library:

>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

>>> tokenizer = AutoTokenizer.from_pretrained("allenai/tk-instruct-3b-def")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("allenai/tk-instruct-3b-def")

>>> input_ids = tokenizer.encode(
        "Definition: return the currency of the given country. Now complete the following example - Input: India. Output:", 
        return_tensors="pt"
    )
>>> output = model.generate(input_ids, max_length=10)
>>> output = tokenizer.decode(output[0], skip_special_tokens=True)

The model should generate 'Indian Rupee' as the output.

Evaluation

The following script evaluates our 3B Tk-Instruct model that uses task definition + 2 positive examples as instructions:

./scripts/eval_tk_instruct.sh

This should give you a ROUGE-L score of ~54.0, as is reported in the Table 3 of our paper.

You can also try other models under different encodings. You can control whether to include definition / explanation, or the number of pos/neg examples, by specifying the arguments in src/run_s2s.py.

The numbers for heuristic baselines and GPT3 can be reproduced by using the following scripts:

./scripts/run_heuristics.sh
./scripts/run_gpt3.sh

Model Predictions and Performance

The predictions of our tested models can be found in the output folder. You can evaluate each predition file in the following way:

python src/compute_metrics.py --predictions output/default/tk-instruct-3b-def-pos/predicted_examples.jsonl --track default --compute_per_category_metrics
python src/compute_metrics.py --predictions output/xlingual/mtk-instruct-3b-def-pos/predicted_examples.jsonl --track xlingual --compute_per_category_metrics

Here are the performance numbers (in ROUGE-L) for our tested models:

Note that these numbers might be different from the numbers reported in the our arxiv paper, because we 1) resampled our evaluation instances; 2) updated our evaluation script. We will update the paper once allowed.

We will keep adding the predictions and performance of new models into this repository.

Citation

@inproceedings{supernaturalinstructions,
  title={Super-NaturalInstructions:Generalization via Declarative Instructions on 1600+ Tasks},
  author={Wang, Yizhong and Mishra, Swaroop and Alipoormolabashi, Pegah and Kordi, Yeganeh and Mirzaei, Amirreza and Arunkumar, Anjana and Ashok, Arjun and Dhanasekaran, Arut Selvan and Naik, Atharva and Stap, David and others},
  booktitle={EMNLP},
  year={2022}
}