added all2023 train-test split with fulltext

README.md

@@ -16,10 +16,30 @@ configs:
   data_files:
   - split: val
     path: "all2023/val.json"
+- config_name: all2023_v2
+  data_files:
+  - split: train
+    path: "all2023_v2/train.json"
+  - split: test
+    path: "all2023_v2/test.json"
 ---
 
 # arXiv Classifier Data
 
+Usage:
+```
+from datasets import load_dataset, DownloadMode
+# download from HuggingFace
+dataset = load_dataset('mlcore/arxiv-classifier', name=<CONFIG NAME>)
+# load from G2
+dataset = load_dataset('/share/nikola/arxiv_classifier/data/arxiv-classifier', name=<CONFIG NAME>)
+```
+To force the dataset to be re-generated:
+```
+dataset = load_dataset('/share/nikola/arxiv_classifier/data/arxiv-classifier', name=<CONFIG NAME>, download_mode=DownloadMode.FORCE_REDOWNLOAD)
+```
+See: https://huggingface.co/docs/datasets/v2.20.0/en/package_reference/builder_classes#datasets.DownloadMode
+
 Standardized terminology:
 - Field: Bio/cs/physics
 - Subfield: Subcategories within each
@@ -38,11 +58,18 @@ Original data: https://www.dropbox.com/scl/fo/wwu0ifghw4sco09g67frb/h?rlkey=6ddg
 **Major**: Dataset of papers between 2010 and 2020 (with some pre-2010 papers) with unbalanced primary subfields to better represent the true distribution of primary categories, which is dominated by a few subfields. Note that the distribution of major subfields is still truncated.
 
 **All 2023**: All papers published in 2023.
+- Train: papers with date between January and June (inclusive)
+- Test: papers with date between July and December (inclusive)
 
 ![Primary subfield distribution](figures/Primary_subfield_distribution.png)
 
 ![Secondary subfield distribution](figures/Secondary_subfield_distribution.png)
 
+To generate subfield distribution plots:
+```
+python plot_subfield_distributions.py --output_path <path_to_save_plots>
+```
+
 ## Set up dependencies
 
 Get subfields to ignore and subfield aliases:

all2023/val.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4186a5287defbe55cb14b48b96a9d42630b623e645f65a65e18bfa12d4f9bfcb
-size 13917099843
+oid sha256:7323614e2e540385689dba0e6252fe16c51e47ad6de946d7c2545f563d4290c3
+size 13921074283

all2023_v2/test.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bf8222a0cbdbc4889512009069292e108477b9ecf5075522a770dcf5137348e
+size 7220291019

all2023_v2/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e06a48f7e9bb778b8333b42481989e2b88a09a9affdbef695bede25c76020c3b
+size 6700804113

figures/primary_subfield_distribution.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:820f2f0256f5ea60b703d2bface4a6933a28366207fbc62b32c878004261fd79
+size 10666

figures/secondary_subfield_distribution.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0da253fe17779bbff8a61789c9a809aa285569754be282c79729585fccd0ed51
+size 10640

plot_subfield_distributions.py

@@ -1,23 +1,45 @@
+# ---
+# jupyter:
+#   jupytext:
+#     cell_metadata_filter: -all
+#     custom_cell_magics: kql
+#     text_representation:
+#       extension: .py
+#       format_name: percent
+#       format_version: '1.3'
+#       jupytext_version: 1.11.2
+#   kernelspec:
+#     display_name: Python 3
+#     language: python
+#     name: python3
+# ---
+
+# %% [markdown]
 # Script to plot the distribution of primary and secondary subfields in the minor, major, and all2023 datasets.
 # Usage:
 #   python plot_subfield_distributions.py --output_path <path_to_save_plots>
 
+# %%
 from argparse import ArgumentParser
-parser = ArgumentParser()
+# parser = ArgumentParser()
 # parser.add_argument("--dataset", "-d", type=str, default="minor")
 # parser.add_argument("--split", "-s", type=str, default="val")
-parser.add_argument("--output_path", "-op", type=str, default="./")
-args = parser.parse_args()
-output_path = args.output_path
+# parser.add_argument("--output_path", "-op", type=str, default="./")
+# args = parser.parse_args()
+# output_path = args.output_path
+output_path = './'
 
+# %%
 # 
 # Remaining imports
 # 
 import os
+import pandas as pd
 import plotly.graph_objects as go
 from datasets import load_dataset
 import itertools
-
+from datetime import datetime
+# %%
 # Get save path
 save_path = os.path.join(
     output_path,
@@ -26,11 +48,13 @@ save_path = os.path.join(
 os.makedirs(save_path, exist_ok=True)
 print(f"Saving figures to {save_path}")
 
+# %%
 # Loading dataset and tokenizer as well
 minor_dataset = load_dataset("mlcore/arxiv-classifier")
 major_dataset = load_dataset("mlcore/arxiv-classifier", "major")
 all2023_dataset = load_dataset("mlcore/arxiv-classifier", "all2023")
 
+# %%
 # Extracting the primary and secondary subfield distributions
 datasets = {
     "minor dataset" : minor_dataset['train'], 
@@ -43,6 +67,7 @@ secondary_subfield_distributions = {
     for name, dataset in datasets.items()
 }
 
+# %%
 # Plotting the two distributions
 hists = []
 n = min([len(i) for i in primary_subfield_distributions.values()])
@@ -50,8 +75,10 @@ for name, primary_subfield_distribution in primary_subfield_distributions.items(
     # primary_subfield_distribution = random.sample(primary_subfield_distribution, n)
     hists.append(go.Histogram(x=primary_subfield_distribution, name=name, opacity=0.75, histnorm='probability'))
 
+# %%
 fig = go.Figure(data=hists)
 
+# %%
 # Update layout for better readability
 fig.update_layout(
     height=700,
@@ -71,11 +98,29 @@ fig.update_layout(
         borderwidth=1
     )
 )
-    
+
+# %%
 # Show the figure
 # fig.show()
 fig.write_image(os.path.join(save_path,"Primary_subfield_distribution.png"))
+# also save distributions as csv
+df_list = []
+for name, primary_subfield_distribution in primary_subfield_distributions.items():
+    df = pd.DataFrame(data={
+        "primary_subfield": primary_subfield_distribution,
+        "dataset": name,
+        "count": 1
+    })
+    df_list.append(df)
+df = pd.concat(df_list)
+counts_by_subfield = df.groupby(['dataset', 'primary_subfield']).agg('count').unstack(level=0)
+counts_by_subfield.columns = counts_by_subfield.columns.droplevel()
+# normalize counts
+counts_by_subfield = counts_by_subfield.div(counts_by_subfield.sum(axis=0), axis=1)
+# print(counts_by_subfield)
+counts_by_subfield.to_csv(os.path.join(save_path, "primary_subfield_distribution.csv"))
 
+# %%
 # Plotting the two distributions
 hists = []
 n = min([len(i) for i in secondary_subfield_distributions.values() if len(i) > 0])
@@ -83,8 +128,10 @@ for name, secondary_subfield_distribution in secondary_subfield_distributions.it
     # secondary_subfield_distribution = random.sample(secondary_subfield_distribution, min(n, len(secondary_subfield_distribution)))
     hists.append(go.Histogram(x=secondary_subfield_distribution, name=name, opacity=0.75, histnorm='probability'))
 
+# %%
 fig = go.Figure(data=hists)
 
+# %%
 # Update layout for better readability
 fig.update_layout(
     height=700,
@@ -104,7 +151,24 @@ fig.update_layout(
         borderwidth=1
     )
 )
-    
+
+# %%
 # Show the figure
 # fig.show()
-fig.write_image(os.path.join(save_path,"Secondary_subfield_distribution.png"))
+fig.write_image(os.path.join(save_path,"Secondary_subfield_distribution.png"))
+# also save distributions as csv
+df_list = []
+for name, secondary_subfield_distribution in secondary_subfield_distributions.items():
+    df = pd.DataFrame(data={
+        "secondary_subfield": secondary_subfield_distribution,
+        "dataset": name,
+        "count": 1
+    })
+    df_list.append(df)
+df = pd.concat(df_list)
+counts_by_subfield = df.groupby(['dataset', 'secondary_subfield']).agg('count').unstack(level=0)
+counts_by_subfield.columns = counts_by_subfield.columns.droplevel()
+# normalize counts
+counts_by_subfield = counts_by_subfield.div(counts_by_subfield.sum(axis=0), axis=1)
+# print(counts_by_subfield)
+counts_by_subfield.to_csv(os.path.join(save_path, "secondary_subfield_distribution.csv"))

preprocess_all2023_v2.py

@@ -30,7 +30,7 @@ parser.add_argument("--input_path", "-ip", type=str, default="./data",
 args, opt = parser.parse_known_args()
 
 dataset = 'all2023' # args.dataset
-split = 'val' # args.split
+# split = 'val' # args.split
 output_path = args.output_path
 input_path = args.input_path
 
@@ -42,6 +42,7 @@ import pandas as pd
 import os
 import sys
 from tqdm import tqdm
+from datetime import datetime
 from util_preprocess import preprocess_primary_secondary
 
 # %% [markdown]
@@ -57,7 +58,6 @@ from util_preprocess import preprocess_primary_secondary
 # - Major category = field
 
 # %%
-# file_path = os.path.join('data', 'raw', dataset, f"{split}_{dataset}_cats_full.json")
 file_path = os.path.join(
     input_path,
     'raw',
@@ -66,7 +66,8 @@ file_path = os.path.join(
 )
 save_dir = os.path.join(output_path, dataset)
 os.makedirs(save_dir, exist_ok=True)
-save_path = os.path.join(save_dir, f"{split}.json")
+train_save_path = os.path.join(save_dir, "train.json")
+test_save_path = os.path.join(save_dir, "test.json")
 print("Reading from file: ", file_path)
 
 FEATURES = [
@@ -81,6 +82,17 @@ FEATURES = [
     'field', 
     'fulltext'
 ]
+def get_date(paper):
+    # first get the arxiv id
+    paper_id = paper['paper_id']
+    # yy is the first two characters of the arxiv id
+    yy = paper_id[:2]
+    # mm is the next two characters of the arxiv id
+    mm = paper_id[2:4]
+    # return datatime object
+    date_obj = datetime(year=int(f"20{yy}"), month=int(mm), day=1)
+    # return date as string in the format YYYY-MM-DDTHH:MM:SS
+    return date_obj #.strftime('%Y-%m-%dT%H:%M:%S')
 
 def get_fulltext(row):
     filename = f"{row['paper_id']}v{row['version']}.txt"
@@ -105,6 +117,10 @@ count_hasfulltext = []
 # Ignore all quotes by setting quoting=3 (corresponds to csv.QUOTE_NONE)
 for i, original_df in tqdm(enumerate(pd.read_csv(file_path, sep='\t', chunksize=10, dtype=str, quoting=3))):
     count_orig.append( len(original_df) )
+    # set `yymm` column
+    original_df.loc[:, 'yymm'] = original_df['paper_id'].apply(lambda x: x[:4])
+    # set `created` column using yymm from the paper_id as a proxy
+    original_df.loc[:, 'created'] = original_df.apply(lambda x: get_date(x), axis=1)
     # set non-existent features to empty string
     for feature in FEATURES:
         if feature not in original_df.columns:
@@ -113,7 +129,7 @@ for i, original_df in tqdm(enumerate(pd.read_csv(file_path, sep='\t', chunksize=
     original_df.loc[:, 'primary_subfield'] = original_df['categories'].apply(lambda x : x.split()[0])
     # keep the secondary_subfield a string, will be converted to list in preprocess_primary_secondary()
     original_df.loc[:, 'secondary_subfield'] = original_df['categories']
-    # Apply preprocessing rules
+    # Apply preprocessing rules and add `field` column
     df = preprocess_primary_secondary(original_df)
 
     count_isvalid.append( len(df) )
@@ -125,13 +141,23 @@ for i, original_df in tqdm(enumerate(pd.read_csv(file_path, sep='\t', chunksize=
     df = df.dropna(subset=['fulltext'])
     count_hasfulltext.append( len(df) )
 
+    # assert that df has all the required columns
+    assert all([feature in df.columns for feature in FEATURES]), f"Missing features in df: {FEATURES}"
+
     # convert all columns to string per huggingface requirements
+    DATE_CUTOFF = datetime(2023, 7, 1)
     if i == 0:
-        df.to_json(save_path, lines=True, orient="records")
+        # save the rows with created date before 2023-06-01 to train.json
+        df[df['created'] < DATE_CUTOFF].to_json(train_save_path, lines=True, orient="records") #, date_format='iso')
+        # save the rows with created date after 2023-06-01 to test.json
+        df[df['created'] >= DATE_CUTOFF].to_json(test_save_path, lines=True, orient="records") #, date_format='iso')
     else:
-        df.to_json(save_path, lines=True, orient="records", mode='a')
+        # save the rows with created date before 2023-06-01 to train.json
+        df[df['created'] < DATE_CUTOFF].to_json(train_save_path, lines=True, orient="records", mode='a')#, date_format='iso')
+        # save the rows with created date after 2023-06-01 to test.json
+        df[df['created'] >= DATE_CUTOFF].to_json(test_save_path, lines=True, orient="records", mode='a')#, date_format='iso')
 
-print("Saved to: ", save_path)
+# print("Saved to: ", save_path)
 df_err = pd.DataFrame({
     'orig': count_orig,
     'isvalid': count_isvalid,

util_preprocess.py

@@ -6,6 +6,17 @@ from arxiv_classifier.utils.maps import (CATEGORY_ALIASES,
 IGNORE = GENERAL_CATEGORIES | IGNORED_CATEGOREIES
 
 def preprocess_primary_secondary(df):
+    """
+    Applies substitutions and filtering to primary and secondary subfields in the DataFrame.
+    Adds a new column `field` to the DataFrame based on the primary subfield.
+    (If there exists a `field` column in the original DataFrame, it will be overwritten.)
+
+    Args:
+        df (pd.DataFrame): DataFrame with columns `primary_subfield` and `secondary_subfield`
+
+    Returns:
+        pd.DataFrame: DataFrame with columns `primary_subfield`, `secondary_subfield`, and `field`
+    """
     def substitute_aliases(s):
         if s in CATEGORY_ALIASES:
             return CATEGORY_ALIASES[s]
@@ -38,6 +49,8 @@ def preprocess_primary_secondary(df):
     # Preprocess secondary subfields
     df.loc[:, 'secondary_subfield'] = df['secondary_subfield'].apply(preprocess_secondary)
 
+    # Generate new column `field` from primary subfield
+    df.loc[:, 'field'] = df['primary_subfield'].apply(lambda x: x.split('.')[0])
     # tests
     # assert that each primary subfield is in SUBFIELD_MAP
     assert all(df['primary_subfield'].isin(SUBFIELD_MAP.keys())), df[~df['primary_subfield'].isin(SUBFIELD_MAP.keys())]