Feature/qualitative analysis notebook

notebooks/mt_browser.py

@@ -0,0 +1,166 @@
+import marimo
+
+__generated_with = "0.22.0"
+app = marimo.App(width="medium")
+
+
+@app.cell
+def _():
+    import marimo as mo
+
+    return (mo,)
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    # Phase 1 Machine Translation Browser
+
+    Browse Phase 1 machine translations side-by-side with source, reference, and evaluation
+    scores (chrF, COMET, CometKiwi). Translations and backtranslations are shown in separate
+    sections.
+    """)
+    return
+
+
+@app.cell
+def _():
+    import pathlib
+
+    import polars as pl
+
+    return pathlib, pl
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    ## Configuration
+
+    Set `DATA_DIR` to the phase-1 data directory (e.g. the project drive or TigerData mount).
+    All data files are resolved relative to this path.
+    """)
+    return
+
+
+@app.cell
+def _(pathlib):
+    # Set this to the phase-1 data directory on the project drive / TigerData mount.
+    # Defaults to the local data directory for local development.
+    DATA_DIR = pathlib.Path(__file__).parent.parent / "data" / "Phase 1"
+    return (DATA_DIR,)
+
+
+@app.cell
+def _(DATA_DIR, pl):
+    # Load full sentence translations and join with eval scores.
+    # Uses the complete notion-sents corpus (all Notion concepts), not the annotation subset.
+    _sents_meta = pl.concat(
+        [
+            pl.read_ndjson(DATA_DIR / f"notion-sents/mt-sents-{m}.jsonl")
+            for m in ["google_tllm", "hymt", "gemma"]
+        ]
+    )
+    _sents_scores = pl.concat(
+        [
+            pl.read_csv(DATA_DIR / f"notion-sents/eval-sents-{m}.csv")
+            for m in ["google_tllm", "hymt", "gemma"]
+        ]
+    )
+    sents_df = _sents_meta.join(_sents_scores, on="tr_id", how="left").with_columns(
+        pl.col("chrf").round(3),
+        pl.col("comet").round(3),
+        pl.col("cometkiwi").round(3),
+    )
+    return (sents_df,)
+
+
+@app.cell
+def _(DATA_DIR, pl):
+    # Load paragraph translations and join with eval scores.
+    _pars_meta = pl.concat(
+        [
+            pl.read_ndjson(DATA_DIR / f"mto-pars/mt-pars-{m}.jsonl")
+            for m in ["google_tllm", "hymt", "gemma"]
+        ]
+    )
+    _pars_scores = pl.concat(
+        [
+            pl.read_csv(DATA_DIR / f"mto-pars/eval-pars-{m}.csv")
+            for m in ["google_tllm", "hymt", "gemma"]
+        ]
+    )
+    pars_df = _pars_meta.join(_pars_scores, on="tr_id", how="left").with_columns(
+        pl.col("chrf").round(3),
+        pl.col("comet").round(3),
+        pl.col("cometkiwi").round(3),
+    )
+    return (pars_df,)
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    ## 1. Sentence translations
+    """)
+    return
+
+
+@app.cell
+def _(pl, sents_df):
+    sents_df.filter(pl.col("src_lang") != "en").select(
+        ["pair_id", "model", "src_lang", "tr_lang", "chrf", "comet", "cometkiwi", "src_text", "ref_text", "tr_text"]
+    )
+    return
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    ## 2. Sentence backtranslations
+    """)
+    return
+
+
+@app.cell
+def _(pl, sents_df):
+    sents_df.filter(pl.col("src_lang") == "en").select(
+        ["pair_id", "model", "src_lang", "tr_lang", "chrf", "comet", "cometkiwi", "src_text", "ref_text", "tr_text"]
+    )
+    return
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    ## 3. Paragraph translations
+    """)
+    return
+
+
+@app.cell
+def _(pars_df, pl):
+    pars_df.filter(pl.col("src_lang") != "en").select(
+        ["pair_id", "model", "src_lang", "tr_lang", "chrf", "comet", "cometkiwi", "src_text", "ref_text", "tr_text"]
+    )
+    return
+
+
+@app.cell
+def _(mo):
+    mo.md("""
+    ## 4. Paragraph backtranslations
+    """)
+    return
+
+
+@app.cell
+def _(pars_df, pl):
+    pars_df.filter(pl.col("src_lang") == "en").select(
+        ["pair_id", "model", "src_lang", "tr_lang", "chrf", "comet", "cometkiwi", "src_text", "ref_text", "tr_text"]
+    )
+    return
+
+
+if __name__ == "__main__":
+    app.run()

pyproject.toml

@@ -23,7 +23,7 @@ classifiers = [
 ]
 dependencies = [
   "pip", # required for downloading spaCy model in uv-created virtualenvs
-  "marimo>=0.13.11",
+  "marimo>=0.22.0",
   "polars>=0.20.4",
   "numpy",
   "ipython", # Required by transformers for trainer functionality
@@ -56,6 +56,8 @@ path = "src/muse/__init__.py"
 target-version = "py312"
 # Configure src path so ruff import fixes can identify local imports
 src = [ "src" ]
+# Ignore notebooks — marimo notebooks are not linted
+extend-exclude = [ "notebooks/*.py" ]
 
 [tool.ruff.lint]
 # Include these rules in addition to ruff's defaults