metadata

language:
  - en
tags:
  - table-to-text
  - tabular
  - Narratable
datasets:
  - totto
widget:
  - text: >+
      <s><page_title> John Higgins </page_title> <section_title> Minor-ranking
      finals: 6 (3 titles, 3 runners-up) </section_title> <table> <row> <c>
      Outcome </c> <c> No. <row_header> Outcome </row_header> </c> <c> Year
      <row_header> Outcome </row_header> <row_header> No. </row_header> </c> <c>
      Championship <row_header> Outcome </row_header> <row_header> No.
      </row_header> <row_header> Year </row_header> </c> <c> Opponent in the
      final <row_header> Outcome </row_header> <row_header> No. </row_header>
      <row_header> Year </row_header> <row_header> Championship </row_header>
      </c> <c> Score <row_header> Outcome </row_header> <row_header> No.
      </row_header> <row_header> Year </row_header> <row_header> Championship
      </row_header> <row_header> Opponent in the final </row_header> </c> </row>
      <row> <c> Winner <col_header> Outcome </col_header> </c> <c> 1.
      <col_header> No. </col_header> </c> <c> 2010 <col_header> Year
      </col_header> </c> <c> Ruhr Championship <col_header> Championship
      </col_header> </c> <c> England Shaun Murphy <col_header> Opponent in the
      final </col_header> </c> <c> 4–2 <col_header> Score </col_header> </c>
      </row> <row> <c> Runner-up <col_header> Outcome </col_header> </c> <c> 1.
      <col_header> No. </col_header> </c> <c> 2010 <col_header> Year
      </col_header> </c> <c> Prague Classic <col_header> Championship
      </col_header> </c> <c> England Michael Holt <col_header> Opponent in the
      final </col_header> </c> <c> 3–4 <col_header> Score </col_header> </c>
      </row> <row> <c> Runner-up <col_header> Outcome </col_header> </c> <c> 2.
      <col_header> No. </col_header> </c> <c> 2011 <col_header> Year
      </col_header> </c> <c> Players Tour Championship – Event 5 <col_header>
      Championship </col_header> </c> <c> England Andrew Higginson <col_header>
      Opponent in the final </col_header> </c> <c> 1–4 <col_header> Score
      </col_header> </c> </row> <row> <c> Winner <col_header> Outcome
      </col_header> </c> <c> 2. <col_header> No. </col_header> </c> <c> 2012
      <col_header> Year </col_header> </c> <c> Kay Suzanne Memorial Trophy
      <col_header> Championship </col_header> </c> <c> England Judd Trump
      <col_header> Opponent in the final </col_header> </c> <c> 4–2 <col_header>
      Score </col_header> </c> </row> <row> <c> Runner-up <col_header> Outcome
      </col_header> </c> <c> 3. <col_header> No. </col_header> </c> <c> 2012
      <col_header> Year </col_header> </c> <c> Bulgarian Open <col_header>
      Championship </col_header> </c> <c> England Judd Trump <col_header>
      Opponent in the final </col_header> </c> <c> 0–4 <col_header> Score
      </col_header> </c> </row> <row> <highlighted_cell> Winner <col_header>
      Outcome </col_header> </highlighted_cell> <c> 3. <col_header> No.
      </col_header> </c> <highlighted_cell> 2013 <col_header> Year </col_header>
      </highlighted_cell> <highlighted_cell> Bulgarian Open <col_header>
      Championship </col_header> </highlighted_cell> <highlighted_cell>
      Australia Neil Robertson <col_header> Opponent in the final </col_header>
      </highlighted_cell> <highlighted_cell> 4–1 <col_header> Score
      </col_header> </highlighted_cell> </row> </table>

inference:
  parameters:
    max_length: 500

BLOOM (0.56B) fine-tuned on ToTTo for Table-to-text 📋 ➡️ 🔤 aka NARRATABLE

This model is a fine-tuned version of bigscience/bloom-560m on the ToTTo dataset.

The model 🧠

It is a 560M params version of BLOOM 🌸

The dataset 📚

ToTTo is an open-domain English table-to-text dataset with over 120,000 training examples that proposes a controlled generation task: given a Wikipedia table and a set of highlighted table cells, produce a one-sentence description.

During the dataset creation process, tables from English Wikipedia are matched with (noisy) descriptions. Each table cell mentioned in the description is highlighted and the descriptions are iteratively cleaned and corrected to faithfully reflect the content of the highlighted cells.

Evaluation results

Metric	Value
rouge1	0.56
rouge2	0.33
rougeL	0.48
rougeLsum	0.48
sacrebleu	20.87
meteor	0.49

Usage

from datasets import load_dataset
from transformers import BloomTokenizerFast, BloomForCausalLM

valid_dataset = load_dataset('totto', split='validation')

from preprocess import preprocess # This file is included in the repo

# Now we linearize the tables
valid_dataset = valid_dataset.map(preprocess) 

model_ckpt = "mrm8488/bloom-560m-finetuned-totto-table-to-text"

tokenizer = BloomTokenizerFast.from_pretrained(ckpt)
model = BloomForCausalLM.from_pretrained(ckpt).to("cuda")


def explain_hl_cells(text):
    inputs = tokenizer(text, return_tensors='pt')
    input_ids = inputs.input_ids.to("cuda")
    attention_mask = inputs.attention_mask.to("cuda")
    output = model.generate(input_ids, attention_mask=attention_mask, max_length=2048, eos_token_id=tokenizer.eos_token_id)

    return tokenizer.decode(output[0], skip_special_tokens=False)

example = valid_dataset[1]

print(explain_hl_cells(example['linearized_table'])

Framework versions

Transformers 4.21.2
Pytorch 1.12.1+cu113
Datasets 2.4.0
Tokenizers 0.12.1

Created by: Narrativa

About Narrativa:

Narrativa is an internationally recognized content services company that uses its proprietary artificial intelligence and machine learning platforms to build and deploy digital content solutions for enterprises. Its technology suite, consisting of data extraction, data analysis, natural language processing (NLP) and natural language generation (NLG) tools, all seamlessly work together to power a lineup of smart content creation, automated business intelligence reporting and process optimization products for a variety of industries. Contact us to learn more about our solutions!