fineweb2-bagaco / scripts /marimo_explore.py
duarteocarmo
update
1eb05b5
Raw
History Blame Contribute Delete
12.8 kB
# /// script
# dependencies = [
# "diskcache==5.6.3",
# "duckdb==1.4.4",
# "marimo",
# "matplotlib==3.10.8",
# "numpy",
# "openai",
# "pandas==3.0.0",
# "polars[pyarrow]==1.38.1",
# "pydantic-ai==1.59.0",
# "scikit-learn==1.8.0",
# "sqlglot==28.10.1",
# "tqdm",
# ]
# requires-python = ">=3.14"
# ///
import marimo
__generated_with = "0.20.1"
app = marimo.App()
@app.cell
def _():
import duckdb
import marimo as mo
import matplotlib.pyplot as plt
import numpy
import polars as pl
conn = duckdb.connect(database="bagaco.duckdb")
return conn, mo, numpy, pl, plt
@app.cell
def _(bagaco, conn, mo):
_df = mo.sql(
f"""
SELECT COUNT(*) AS total_docs FROM bagaco;
""",
engine=conn,
)
mo.vstack(items=[mo.md(text="### Total document count"), _df])
return
@app.cell
def _(bagaco, conn, mo):
_df = mo.sql(
f"""
SELECT
regexp_replace(lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)), '^www\\.', '') AS domain,
COUNT(*) AS n_docs
FROM bagaco
WHERE url IS NOT NULL AND url <> ''
GROUP BY 1
ORDER BY n_docs DESC
LIMIT 10;
""",
engine=conn,
)
mo.vstack(items=[mo.md(text="### Top 10 domains by document count"), _df])
return
@app.cell
def _(bagaco, conn, mo):
_df = mo.sql(
f"""
SELECT
regexp_replace(lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)), '^www\\.', '') AS domain,
ROUND(AVG(educational_score), 2) AS avg_edu_score,
COUNT(*) AS n_docs
FROM bagaco
WHERE url IS NOT NULL AND url <> '' AND educational_score IS NOT NULL
GROUP BY 1
HAVING COUNT(*) >= 100
ORDER BY avg_edu_score DESC
LIMIT 15;
""",
engine=conn,
)
mo.vstack(
items=[
mo.md(
text="### Domains with highest avg educational score (min 100 docs)"
),
_df,
]
)
return
@app.cell
def _(bagaco, conn, mo):
_df = mo.sql(
f"""
SELECT
regexp_replace(lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)), '^www\\.', '') AS domain,
ROUND(AVG(educational_score), 2) AS avg_edu_score,
COUNT(*) AS n_docs
FROM bagaco
WHERE url IS NOT NULL AND url <> '' AND educational_score IS NOT NULL
GROUP BY 1
HAVING COUNT(*) >= 100
ORDER BY avg_edu_score ASC
LIMIT 15;
""",
engine=conn,
)
mo.vstack(
items=[
mo.md(
text="### Domains with lowest avg educational score (min 100 docs)"
),
_df,
]
)
return
@app.cell
def _(bagaco, conn, mo, plt):
_domain_scatter_df = mo.sql(
f"""
WITH domain_stats AS (
SELECT
regexp_replace(lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)), '^www\\.', '') AS domain,
COUNT(*) AS total_docs,
AVG(educational_score) AS avg_edu_score
FROM bagaco
WHERE url IS NOT NULL
AND url <> ''
AND educational_score IS NOT NULL
GROUP BY 1
HAVING COUNT(*) > 10
),
domain_categories AS (
SELECT
domain,
category,
ROW_NUMBER() OVER (PARTITION BY domain ORDER BY n_docs DESC, category ASC) AS row_number
FROM (
SELECT
regexp_replace(lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)), '^www\\.', '') AS domain,
category,
COUNT(*) AS n_docs
FROM bagaco
WHERE url IS NOT NULL
AND url <> ''
AND category IS NOT NULL
GROUP BY 1, 2
)
)
SELECT
domain_stats.domain,
domain_stats.total_docs,
ROUND(domain_stats.avg_edu_score, 3) AS avg_edu_score,
COALESCE(domain_categories.category, 'Unknown') AS category
FROM domain_stats
LEFT JOIN domain_categories
ON domain_stats.domain = domain_categories.domain
AND domain_categories.row_number = 1
ORDER BY domain_stats.total_docs DESC;
""",
engine=conn,
)
_categories = sorted(_domain_scatter_df["category"].unique().to_list())
_colors = plt.cm.tab10.colors
_category_colors = {
_category: _colors[_index % len(_colors)]
for _index, _category in enumerate(_categories)
}
_, _ax = plt.subplots(figsize=(12, 6))
for _category in _categories:
_subset = _domain_scatter_df.filter(
_domain_scatter_df["category"] == _category
)
_ax.scatter(
x=_subset["total_docs"].to_list(),
y=_subset["avg_edu_score"].to_list(),
s=16,
alpha=0.65,
color=_category_colors[_category],
label=_category,
)
_ax.set_xscale(value="log")
_ax.set_xlabel(xlabel="Total documents (log scale)")
_ax.set_ylabel(ylabel="Average educational score")
_ax.set_title(
label="Domain Scatter: Total Documents (Log) vs Avg Educational Score",
fontweight="bold",
)
_ax.grid(visible=True, axis="both", linestyle="-", linewidth=0.5, alpha=0.35)
_ax.legend(loc="best", fontsize=8, ncols=2)
plt.tight_layout()
_ax
return
@app.cell
def _(bagaco, conn, mo):
_fast_total_words = mo.sql(
f"""
SELECT
SUM(
CASE
WHEN text IS NULL OR text = '' THEN 0
ELSE len(trim(text)) - len(replace(trim(text), ' ', '')) + 1
END
) AS total_words_fast
FROM bagaco;
""",
engine=conn,
)
mo.vstack(
items=[
mo.md(text="### Total words in dataset (fast computation)"),
_fast_total_words,
]
)
return
@app.cell
def _(bagaco, conn, mo, plt):
df_per_halfyear = mo.sql(
f"""
SELECT
EXTRACT(YEAR FROM CAST(date AS TIMESTAMP)) || ' H' ||
CASE WHEN EXTRACT(MONTH FROM CAST(date AS TIMESTAMP)) <= 6 THEN 1 ELSE 2 END AS year_half,
COUNT(*) / 1000000.0 AS total_doc_count_millions,
ROUND(AVG(educational_score), 3) AS avg_edu_score
FROM bagaco
WHERE date IS NOT NULL
GROUP BY 1
ORDER BY 1;
""",
engine=conn,
)
year_half = df_per_halfyear["year_half"].to_list()
total_doc_count_millions = df_per_halfyear[
"total_doc_count_millions"
].to_list()
_avg_edu = df_per_halfyear["avg_edu_score"].to_list()
_x_positions = list(range(len(year_half)))
_fig, ax = plt.subplots(figsize=(12, 6))
ax.bar(
x=_x_positions,
height=total_doc_count_millions,
width=0.65,
color="#449DE3",
edgecolor="#4a4a4a",
linewidth=0.6,
alpha=0.85,
label="Total Document Count (M)",
)
ax.set_xticks(ticks=_x_positions)
ax.set_xticklabels(labels=year_half, rotation=45, ha="right")
ax.set_xlabel(xlabel="Half-Year")
ax.set_ylabel(ylabel="Total Document Count (Millions)")
ax.grid(axis="y", linestyle="-", linewidth=0.5, alpha=0.35)
_ax2 = ax.twinx()
_ax2.plot(
_x_positions,
_avg_edu,
color="#d73027",
linewidth=2,
marker="o",
markersize=4,
label="Avg Edu Score",
)
_ax2.set_ylabel(ylabel="Avg Educational Score", color="#d73027")
_ax2.tick_params(axis="y", labelcolor="#d73027")
_lines1, _labels1 = ax.get_legend_handles_labels()
_lines2, _labels2 = _ax2.get_legend_handles_labels()
ax.legend(
handles=_lines1 + _lines2, labels=_labels1 + _labels2, loc="upper left"
)
ax.set_title(
label="Document Volume & Avg Educational Score Over Time",
fontweight="bold",
)
plt.tight_layout()
_fig
return
@app.cell
def _(bagaco, conn, mo, plt):
df_word_count_distribution = mo.sql(
f"""
WITH wc AS (
SELECT len(string_split(text, ' ')) AS word_count FROM bagaco WHERE text IS NOT NULL
), bucketed AS (
SELECT
CASE
WHEN word_count <= 150 THEN 'Note (0-150)'
WHEN word_count <= 600 THEN 'Short Article (151-600)'
WHEN word_count <= 1500 THEN 'Standard Article (601-1500)'
WHEN word_count <= 5000 THEN 'Longform (1501-5000)'
WHEN word_count <= 20000 THEN 'Deep Report (5001-20000)'
ELSE 'Book (20001+)'
END AS length_band,
COUNT(*) / 1000000.0 AS doc_count_millions
FROM wc
GROUP BY 1
)
SELECT
CASE
WHEN length_band = 'Note (0-150)' THEN 1
WHEN length_band = 'Short Article (151-600)' THEN 2
WHEN length_band = 'Standard Article (601-1500)' THEN 3
WHEN length_band = 'Longform (1501-5000)' THEN 4
WHEN length_band = 'Deep Report (5001-20000)' THEN 5
ELSE 6
END AS bucket_order,
length_band,
doc_count_millions
FROM bucketed
ORDER BY 1;
""",
engine=conn,
)
length_band_labels = df_word_count_distribution["length_band"].to_list()
doc_count_millions = df_word_count_distribution["doc_count_millions"].to_list()
y_positions = list(range(len(length_band_labels)))
_, word_count_ax = plt.subplots(figsize=(12, 4.8))
word_count_ax.barh(
y=y_positions,
width=doc_count_millions,
color="#449DE3",
edgecolor="#4a4a4a",
linewidth=0.6,
alpha=0.85,
label="Total Document Count (M)",
)
word_count_ax.set_yticks(ticks=y_positions)
word_count_ax.set_yticklabels(labels=length_band_labels)
word_count_ax.set_xlabel(xlabel="Total Document Count (Millions)")
word_count_ax.set_ylabel(ylabel="Word Count Band")
word_count_ax.set_title(label="Word Count Distribution", fontweight="bold")
word_count_ax.grid(
visible=True, axis="x", linestyle="-", linewidth=0.5, alpha=0.35
)
word_count_ax.legend(loc="lower right")
plt.tight_layout()
word_count_ax
return
@app.cell
def _(bagaco, conn, mo, numpy, pl, plt):
_df_heatmap = mo.sql(
f"""
SELECT
category,
EXTRACT(YEAR FROM CAST(date AS TIMESTAMP)) || ' H' ||
CASE WHEN EXTRACT(MONTH FROM CAST(date AS TIMESTAMP)) <= 6 THEN 1 ELSE 2 END AS year_half,
ROUND(AVG(educational_score), 3) AS avg_edu_score
FROM bagaco
WHERE date IS NOT NULL AND category IS NOT NULL AND educational_score IS NOT NULL
GROUP BY 1, 2
ORDER BY 1, 2;
""",
engine=conn,
)
_categories = sorted(_df_heatmap["category"].unique().to_list())
_periods = sorted(_df_heatmap["year_half"].unique().to_list())
_grid = numpy.full(
shape=(len(_categories), len(_periods)), fill_value=numpy.nan
)
for _i, _cat in enumerate(_categories):
for _j, _period in enumerate(_periods):
_match = _df_heatmap.filter(
(pl.col("category") == _cat) & (pl.col("year_half") == _period)
)
if len(_match) > 0:
_grid[_i, _j] = _match["avg_edu_score"][0]
_fig, _ax = plt.subplots(figsize=(14, 6))
_im = _ax.imshow(_grid, aspect="auto", cmap="RdYlGn", interpolation="nearest")
plt.colorbar(mappable=_im, ax=_ax, label="Avg Educational Score")
_ax.set_yticks(ticks=range(len(_categories)))
_ax.set_yticklabels(labels=_categories)
_ax.set_xticks(ticks=range(len(_periods)))
_ax.set_xticklabels(labels=_periods, rotation=45, ha="right")
_ax.set_title(
label="Avg Educational Score by Category and Half-Year", fontweight="bold"
)
plt.tight_layout()
_fig
return
@app.cell
def _(bagaco, conn, mo):
_category_summary_df = mo.sql(
f"""
SELECT
category,
COUNT(*) AS total_documents,
ROUND(AVG(educational_score), 3) AS average_educational_score,
ROUND(AVG(educational_score), 3) AS mean_educational_score,
ROUND(
SUM(CASE WHEN educational_score >= 3 THEN 1 ELSE 0 END) * 100.0 / COUNT(*),
2
) AS pct_high_educational_score_documents,
COUNT(
DISTINCT regexp_replace(
lower(regexp_extract(url, '^(?:https?://)?(?:[^@/\\n]+@)?([^:/?\\n]+)', 1)),
'^www\\.',
''
)
) AS total_different_domains
FROM bagaco
WHERE category IS NOT NULL
AND educational_score IS NOT NULL
GROUP BY 1
ORDER BY total_documents DESC;
""",
engine=conn,
)
mo.vstack(items=[mo.md(text="### Category summary"), _category_summary_df])
return
@app.cell
def _():
return
if __name__ == "__main__":
app.run()