Skip to content

Latest commit

 

History

History
68 lines (36 loc) · 3.81 KB

README.md

File metadata and controls

68 lines (36 loc) · 3.81 KB

EDGAR10-Q Dataset

This dataset is built from 10-Q/K documents (Quarterly and Yearly Reports) of publicly listed companies on the SEC. To access these documents, follow this link. Please see sample.csv to find the instance of a document of the dataset. To get CIK of an organization, use the CIK_lookup in contents folder.

Data Fields

The data fields are the same among all splits.

  • text: a string in the form of entity plus sentence.
  • label: a string describing the relevant context for entity in the sentence

Data Splits

The dataset is split into train, validation, and test sets. The sizes of the splits are as follows:

Train Validation Test
Instances 1,498,995 187,383 187,383

Building Dataset

Using the script dataset_generation_and_baseline.py will pull the the data from sec website and store it in content folder. cik_lookup.xlsx has the list of 2000 organizations whose data was pulled. The script will also run the baseline approach and store all the results in each organizations' excel respectively.

ChatGPT response generation

Once the dataset is created and baseline appraoch is executed and the excel is complete, use the script chatgpt_responses.py for getting the reuslts from ChatGPT. Please use your own API key for its execution.

Table 1 : Instance of the Dataset.

Sentence value entity type Labels for each entity
As of August 5, 2019, there were 46,662,179 shares of common stock, $0.01 par value, outstanding. 4,66,62,179 CARDINAL Shares Outstanding
The Company also derecognized existing deferred rent liabilities of $15,302. 15,302 MONEY Rent Expense
The intangible assets acquired have a weighted average useful life of approximately nine years. nine years DATE Intangible assets
The initial purchase price of $31,676 included $30,176 cash consideration paid upon acquisition, funded primarily through borrowings under the Senior 31,676 MONEY Initial purchase price
Credit Facility, and a contingent earn out payment of up to $25,000 with an estimated fair value of $1,500 as of the acquisition date. 30,176 MONEY Payments to Acquire Businesses

Table 2: Statistics about dataset:

Imgur

Results for baseline algorithm, ChatGPT responses and supervised learning models

Imgur

Supervised Finetuning

Use supervised_deepspeed_finetuning.sh for finetuning any model on EDGAR10-Q dataset.

Results on Dowstream datasets

[EDGAR-T5-Large] was finetuned on some downstream datasets to get better results than T5 large. BloombergGPT 50B was used as baseline.

Dataset Bloomberg GPT 50B T5 Large Edgar T5 Large
FiQA SA 75.07 74.89 80.42
FPB 51.07 55.77 79.69
Headline 82.20 90.55 93.55