SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2111.10952 · ICLR · 2022

ExT5: Towards Extreme Multi-Task Scaling for Transfer Learning

Yi Tay, Donald Metzler, Tal Schuster, Jinfeng Rao, Vamsi Aribandi, Steven Zheng, Vaibhav Sanket, Honglei Mehta, Vinh Zhuang, Dara Tran, Jianmo Bahri, Jai Ni, and 3 more

arXiv · PDF · Open in the Atlas

Code that ran

6 of 9 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 6 have no contract verdict; 0 failures of our instrument; 3 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2111.10952.json · badge: /api/ran/2111.10952.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
google-research/text-to-text-transfer-transformer canonical (classifier label)
classifier basis not recorded on the link
6 ran · 3 no recorded run
FunctionStatusWhere it lives
filter_features Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/models/utils.py
code served (permissive licence) · get_code("ea645f2c33765137")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct an example dictionary with all model features and some extra features
ex = {
    'inputs': torch.randn(2, 4),
    'inputs_position': torch.randn(2, 3),
    'inputs_segmentation': torch.randn(2, 5),
    'targets': torch.randn(2, 4),
    'targets_position': torch.randn(2, 3),
    'targets_segmentation': torch.randn(2, 5),
    'targets_subsegmentation': torch.randn(2, 6),
    'extra_feature_1': torch.randn(2, 2),
    'extra_feature_2': 'should_be_filtered_out',
    'metadata': {'some': 'data'},
}

ARGS = {
    'ex': ex
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_step_from_checkpoint_path Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/models/utils.py
code served (permissive licence) · get_code("a4895b42cf3f388d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def get_step_from_checkpoint_path(checkpoint_path):
    """Returns the global step for the checkpoint at `checkpoint_path`.

  **Note**: This function works for checkpoints that are saved in the TF format
  only.

  Assumes `checkpoint_path` corresponds to a file which contains the substring
  model.ckpt-{global_step}

  Args:
    checkpoint_path: str of path to a checkpoint file.

  Returns:
    int of the global step corresponding to the checkpoint file.

  Raises:
    ValueError if checkpoint_path does not correspond to a model checkpoint file
    which contains the global_step in its filename.
  """
    match = re.match('.*model\\.ckpt\\-(\\d+).*', checkpoint_path)
    if match is None:
        raise ValueError('Invalid checkpoint path {}'.format(checkpoint_path))
    return int(match.group(1))


ARGS = {
    'checkpoint_path': '/path/to/model.ckpt-42'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_squad Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/evaluation/qa_utils.py
code served (permissive licence) · get_code("62dbc58154aa737e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
import string

def _normalize_answer(text, punc_chars, punc_repl):
    """Lower text and remove punctuation, articles and extra whitespace."""

    def remove_articles(s):
        return re.sub('\\b(a|an|the)\\b', ' ', s)

    def replace_punctuation(s):
        to_replace = set(punc_chars)
        return ''.join((punc_repl if ch in to_replace else ch for ch in s))

    def white_space_fix(s):
        return ' '.join(s.split())
    text = text.lower()
    text = replace_punctuation(text)
    text = remove_articles(text)
    text = white_space_fix(text)
    return text

def normalize_squad(answer):
    """Normalization used in official SQuAD evaluation script."""
    return _normalize_answer(answer, punc_chars=string.punctuation, punc_repl='')

ARGS = {
    "answer": "The quick brown fox jumps over the lazy dog!"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
rouge Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/evaluation/metrics.py
code served (permissive licence) · get_code("46c76c8dc7bbf945")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

ARGS = {
    'targets': [
        'This is a test summary . It contains multiple sentences .',
        'Another target text . With some content . To evaluate .',
    ],
    'predictions': [
        'This is a test summary . It has similar content .',
        'Another predicted text . With comparable content . For evaluation .',
    ],
    'score_keys': ('rouge1', 'rouge2', 'rougeLsum'),
    'verbose': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, numpy, rouge_score, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
rouge_mean Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/evaluation/metrics.py
code served (permissive licence) · get_code("2274158e8735347f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import collections
from rouge_score import rouge_scorer

torch_manual_seed_value = 0

ARGS = {
    'targets': [
        'This is the first target summary. It contains multiple sentences.',
        'This is the second target summary with different content.'
    ],
    'predictions': [
        'This is the first prediction summary. It has similar content.',
        'This is the second prediction with some variations.'
    ],
    'score_keys': ('rouge1', 'rouge2', 'rougeLsum'),
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
rouge_score — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sort_columns Ran
ran on a synthesized input
google-research/text-to-text-transfer-transformer/t5/evaluation/eval_utils.py
code served (permissive licence) · get_code("c4b3db3eab3b63b8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned DataFrame. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import collections
import pandas as pd

class Metric(object):
    def __init__(self, name, group=None):
        self.name = name
        self.group = group or name

METRIC_NAMES = collections.OrderedDict([
    ('glue_average', Metric('Average GLUE Score')),
    ('glue_cola_v002/matthews_corrcoef', Metric('CoLA')),
    ('glue_sst2_v002/accuracy', Metric('SST-2')),
    ('glue_mrpc_v002/f1', Metric('MRPC (F1)', 'MRPC')),
    ('glue_mrpc_v002/accuracy', Metric('MRPC (accuracy)', 'MRPC')),
])

# Create a sample DataFrame with some metric names as columns
df = pd.DataFrame({
    'Average GLUE Score': [0.85],
    'SST-2': [0.92],
    'CoLA': [0.88],
    'Other Column': [0.75],
})

ARGS = {
    'df': df,
    'metric_names': METRIC_NAMES,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_eval_metric_values Not yet run
unverified (harvested, no recorded run)
google-research/text-to-text-transfer-transformer/t5/evaluation/eval_utils.py
code served (permissive licence) · get_code("f9701b35213138f0")
How it ran

no invocation recorded.

normalize_trivia_qa Not yet run
unverified (harvested, no recorded run)
google-research/text-to-text-transfer-transformer/t5/evaluation/qa_utils.py
code served (permissive licence) · get_code("748cf9475c59f2c2")
How it ran

no invocation recorded.

qa_metrics Not yet run
unverified (harvested, no recorded run)
google-research/text-to-text-transfer-transformer/t5/evaluation/qa_utils.py
code served (permissive licence) · get_code("4c75d412b9154a9d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 8 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
sort_columns 1 other paper
1 dated earlier
1910.10683 (2019-10)
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Despite the recent success of multi-task learning and transfer learning for natural language processing (NLP), few works have systematically studied the effect of scaling up the number of tasks during pre-training. Towards this goal, this paper introduces EXMIX (Extreme Mixture): a massive collection of 107 supervised NLP tasks across diverse domains and task-families. Using EXMIX, we study the effect of multi-task pre-training at the largest scale to date, and analyze cotraining transfer amongst common families of tasks. Through this analysis, we show that manually curating an ideal set of tasks for multi-task pre-training is not straightforward, and that multi-task scaling can vastly improve models on its own. Finally, we propose EXT5: a model pre-trained using a multi-task objective of self-supervised span denoising and supervised EXMIX. Via extensive experiments, we show that EXT5 outperforms strong T5 baselines on SuperGLUE, GEM, Rainbow, Closed-Book QA tasks, and several tasks outside of EXMIX. EXT5 also significantly improves sample efficiency while pre-training. * Google AI Resident. † Equal contribution. Sebastian is now at Google Research. Sanket returned to CMU.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2111.10952")
get_code_for_paper("2111.10952")
have("2111.10952")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2111.10952.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2111.10952.svg)](https://syntology.ai/paper/2111.10952)

Connect an agent — have() is free.