README.md

---
license: mit
language:
- en
tags:
- t5
model-index:
- name: metro_t0p_base
  results:
  - task:
      type: natural-language-inference
    dataset:
      type: super_glue
      name: RTE
      config: rte
      split: validation
    metrics:
      - type: accuracy
        value: 64.90974729241879
  - task:
      type: natural-language-inference
    dataset:
      type: super_glue
      name: CB
      config: cb
      split: validation
    metrics:
      - type: accuracy
        value: 44.642857142857146
  - task:
      type: natural-language-inference
    dataset:
      type: anli
      name: ANLI R1
      split: dev_r1
    metrics:
      - type: accuracy
        value: 32.35333333333333
  - task:
      type: natural-language-inference
    dataset:
      type: anli
      name: ANLI R2
      split: dev_r2
    metrics:
      - type: accuracy
        value: 32.199999999999996
  - task:
      type: natural-language-inference
    dataset:
      type: anli
      name: ANLI R3
      split: dev_r3
    metrics:
      - type: accuracy
        value: 32.9
  - task:
      type: coreference-resolution
    dataset:
      type: super_glue
      name: WSC
      config: wsc.fixed
      split: validation
    metrics:
      - type: accuracy
        value: 61.34615384615385
  - task:
      type: coreference-resolution
    dataset:
      type: winogrande
      name: Winogrande XL
      config: winogrande_xl
      split: validation
    metrics:
      - type: accuracy
        value: 50.860299921073405
  - task:
      type: multiple-choice-qa
    dataset:
      type: super_glue
      name: COPA
      config: copa
      split: validation
    metrics:
      - type: accuracy
        value: 61.5
  - task:
      type: multiple-choice-qa
    dataset:
      type: story_cloze
      name: StoryCloze 2016
      config: '2016'
      split: validation
    metrics:
      - type: accuracy
        value: 82.59754142169962
  - task:
      type: multiple-choice-qa
    dataset:
      type: hellaswag
      name: HellaSwag
      split: validation
    metrics:
      - type: accuracy
        value: 43.22097191794464
  - task:
      type: word-sense-disambiguation
    dataset:
      type: super_glue
      name: WiC
      config: wic
      split: validation
    metrics:
      - type: accuracy
        value: 51.20689655172414
---

Official repository: https://github.com/gonglinyuan/metro_t0

# METRO-T0

Paper: [Model-Generated Pretraining Signals Improves Zero-Shot Generalization of Text-to-Text Transformers](https://arxiv.org/abs/2305.12567) (ACL 2023)

METRO-T0 is a T5-style text-to-text Transformer pretrained using model-generated pretraining signals, prompt-finetuned on a family of public NLP tasks proposed in [T0](https://arxiv.org/abs/2110.08207).
METRO-T0 is highly parameter efficient. For example, METRO-T0-Large++ (775M parameters) outperforms GPT-3 (175B parameters) and T0-3B (3B parameters) on a wide range of NLP tasks.

![The architecture of METRO-T0 during pretraining using BERT as the auxiliary model to generate signals](https://github.com/gonglinyuan/metro_t0/raw/main/assets/metro_t0_method.png)

![Prompt learning results of METRO-T0 versus our T0 baseline and T03B by Sanh et al. (2022) on 4 tasks  in the T0 Eval benchmark. Each point denotes the accuracy using one prompt template, except that the median accuracy over all templates of T03B is indicated by the blue point. The plots of other tasks are in our paper.](https://github.com/gonglinyuan/metro_t0/raw/main/assets/metro_t0_selected_results.png)

## Use METRO-T0+-Base

To use METRO-T0+-Base in PyTorch (Python 3.7+, PyTorch 1.12+ and transformers 4.17+ are prerequisites), refer to the code snippet below:

```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("gonglinyuan/metro_t0p_base", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("gonglinyuan/metro_t0p_base", trust_remote_code=True)

input_text = "Is this review positive or negative? Review: this is the best cast iron skillet you will ever buy"
inputs = tokenizer([input_text], max_length=512, truncation=True, add_special_tokens=True, return_tensors="pt").input_ids
outputs = model.generate(inputs, max_new_tokens=256, do_sample=False)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))  # expected: positive
```

## Other METRO-T0 Models

|                    | # Parameters | Pretraining Data | Prompt-Finetuning Data |
|--------------------|--------------|------------------|------------------------|
| [METRO-T0-Base](https://huggingface.co/gonglinyuan/metro_t0_base)      | 226M         | Wikibook (16G)   | T0 Train               |
| [METRO-T0+-Base](https://huggingface.co/gonglinyuan/metro_t0p_base)     | 226M         | Wikibook (16G)   | T0+ Train              |
| [METRO-T0++-Base](https://huggingface.co/gonglinyuan/metro_t0pp_base)    | 226M         | Wikibook (16G)   | T0++ Train             |
| [METRO-T0-Base++](https://huggingface.co/gonglinyuan/metro_t0_basepp)    | 256M         | 160G corpus      | T0 Train               |
| [METRO-T0+-Base++](https://huggingface.co/gonglinyuan/metro_t0p_basepp)   | 256M         | 160G corpus      | T0+ Train              |
| [METRO-T0++-Base++](https://huggingface.co/gonglinyuan/metro_t0pp_basepp)  | 256M         | 160G corpus      | T0++ Train             |
| [METRO-T0-Large++](https://huggingface.co/gonglinyuan/metro_t0_largepp)   | 775M         | 160G corpus      | T0 Train               |
| [METRO-T0+-Large++](https://huggingface.co/gonglinyuan/metro_t0p_largepp)  | 775M         | 160G corpus      | T0+ Train              |
| [METRO-T0++-Large++](https://huggingface.co/gonglinyuan/metro_t0pp_largepp) | 775M         | 160G corpus      | T0++ Train             |


## Citation

If you find the code and models useful for your research, please cite the following paper:

```
@misc{gong2023modelgenerated,
      title={Model-Generated Pretraining Signals Improves Zero-Shot Generalization of Text-to-Text Transformers}, 
      author={Linyuan Gong and Chenyan Xiong and Xiaodong Liu and Payal Bajaj and Yiqing Xie and Alvin Cheung and Jianfeng Gao and Xia Song},
      year={2023},
      eprint={2305.12567},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2305.12567}
}
```