FINAL FINAL UPDATES TO app.py

app.py

@@ -22,7 +22,7 @@ from transformers import AutoTokenizer, AutoModelForSequenceClassification
 from peft import PeftModel
 import shap
 
-#nltk bits used by vader, tf-idf cleaning and the sentence splitter (same as my notebooks)
+#nltk bits used by vader, tf-idf cleaning and the sentence splitter 
 for pkg in ["stopwords", "vader_lexicon", "punkt", "punkt_tab"]:
     nltk.download(pkg, quiet=True)
 
@@ -49,19 +49,19 @@ VADER_COMPOUND, VADER_NEG = -0.30, 0.02
 NO_GATE = {"OUTSIDE_COMMS"}   #outside comms fires on the keyword alone
 sia = SentimentIntensityAnalyzer()
 
-#2. tf-idf + logistic regression (one model per category, saved from my layer 6 notebook)
+#2. tf-idf + logistic regression 
 vec = joblib.load(os.path.join(HERE, "tfidf_vectorizer.pkl"))
 tf_models = joblib.load(os.path.join(HERE, "tfidf_logistic_models.pkl"))
 STOP = set(stopwords.words("english"))
 stemmer = SnowballStemmer("english")
 
-#3. bert-base-uncased + my lora adapter (the adapter files sit next to this app.py)
+#3. bert-base-uncased + my lora adapter
 
 BERT_THRESHOLDS = {"WORKLOAD": 0.65, "TEAM": 0.90, "ADJUSTMENT": 0.85, "OUTSIDE_COMMS": 0.90}
 tok = AutoTokenizer.from_pretrained(HERE)
 base = AutoModelForSequenceClassification.from_pretrained(
     "bert-base-uncased", num_labels=4, problem_type="multi_label_classification")
-bert = PeftModel.from_pretrained(base, HERE, torch_device="cpu").eval()  
+bert = PeftModel.from_pretrained(base, HERE, torch_device="cpu").eval()   #cpu --> zerogpu pretends cuda exists at startup, eval() turns dropout off
 
 
 def vader_flags(seg):
@@ -156,14 +156,14 @@ def diarise(text):
     return "\n\n".join(kept), "fallback", [s for s in speakers if s != astro], speakers, len(kept), len(turns)
 
 
-#segmentation: whole sentences packed up to 300 words and nothing under 20 words
+#segmentation: whole sentences packed up to 300 words, nothing under 20 words
 MIN_WORDS, MAX_WORDS = 20, 300
 
 
 def segment(text):
     blocks = [b.strip() for b in text.split("\n\n") if b.strip()]
     paras = blocks[:1]
-    for b in blocks[1:]:   
+    for b in blocks[1:]:   #a blank line is only a real break if the sentence ended and the next starts with a capital
         if paras[-1][-1] in ".!?\"\u201d" and b[0].isupper():
             paras.append(b)
         else:
@@ -176,7 +176,7 @@ def segment(text):
             chunks.append(p)
             continue
         sents = []
-        for s in sent_tokenize(p):   
+        for s in sent_tokenize(p):   #glue back "sentences" that start lowercase (punkt splits on initials)
             if sents and s[:1].islower():
                 sents[-1] += " " + s
             else:
@@ -226,7 +226,7 @@ def analyse(file, text):
     elif method == "fallback":
         s3 = f"PASSED (fallback): no 'Interviewed by' line, kept the longest speaker. Removed: {', '.join(removed)}."
     else:
-        s3 = "NO SPEAKER TAGS FOUND so the whole text is treated as the astronaut's words."
+        s3 = "NO SPEAKER TAGS FOUND (like ARMSTRONG:), so the whole text is treated as the astronaut's words."
     if method != "no_tags" and not narrative.strip():
         #known failure from my report: a curly apostrophe or accent in the astronaut's tag (O’CONNOR) isn't matched
         s3 = (f"FAILED: interviewer found ({', '.join(removed)}) but nothing was left after removing them, "
@@ -234,7 +234,7 @@ def analyse(file, text):
     s3 += f" Words left: {len(narrative.split()):,}."
     if method == "no_tags":
         #no cover page and no NAME: tags --> probably not an oral history, the models still run but read with care
-        s3 += " Note: this does not look like a NASA oral history transcript !! so the scores below are less meaningful."
+        s3 += " Note: this does not look like a NASA oral history transcript, so the scores below are less meaningful."
 
     #step 4
     segs = segment(narrative)
@@ -321,7 +321,7 @@ def find_files(pairs):
 #  the page 
 
 CSS = """
-.gradio-container {background: radial-gradient(ellipse at top, #141b33 0%, #05070f 70%) !important;}
+.gradio-container {background: radial-gradient(ellipse at top, #2a0f1a 0%, #05070f 70%) !important;}
 * {box-shadow: none !important;}
 .seg {border: 1px solid #ffffff44; border-radius: 12px; padding: 8px 12px; margin: 6px 0; color: #e6ebff;}
 table {width: 100%; border-collapse: collapse; color: #e6ebff;}
@@ -333,7 +333,7 @@ with gr.Blocks(title="SPACEMIND") as demo:
     gr.Markdown("# SPACEMIND")
 
     with gr.Row():
-        #left side: inputs
+        #inputs
         with gr.Column():
             textInput = gr.Textbox(label="Paste a transcript or a passage", lines=10)
             fileInput = gr.File(label="...or upload a .txt / .pdf transcript", file_types=[".txt", ".pdf"])
@@ -345,13 +345,13 @@ with gr.Blocks(title="SPACEMIND") as demo:
             heldOut = find_files(TEST_SET)
             if heldOut:
                 gr.Examples([[p] for p, _ in heldOut], inputs=fileInput, example_labels=[l for _, l in heldOut],
-                            label="Load a transcript from my test set")
+                            label="Load a transcript from my held-out test set")
             tests = find_files(PIPELINE_TESTS)
             if tests:
                 gr.Examples([[p] for p, _ in tests], inputs=fileInput, example_labels=[l for _, l in tests],
-                            label="Pipeline tests")
+                            label="Pipeline tests (where it works and where it breaks)")
 
-        #right side: the stages
+        #the stages
         with gr.Column():
             out1 = gr.Textbox(label="1. Input (word count)")
             out2 = gr.Textbox(label="2. Cleaning")
@@ -372,9 +372,9 @@ with gr.Blocks(title="SPACEMIND") as demo:
     shapButton.click(explain, [state, segPick, catPick], shapOut)
 
 
-   BURGUNDY = gr.themes.Color(c50="#fdf2f4", c100="#fbe5e9", c200="#f5c6d0", c300="#ec9aab", c400="#d9637c",
-                              c500="#9b2743", c600="#801f38", c700="#661a2e", c800="#4f1524", c900="#3d1019",
-                              c950="#2e0a14", name="burgundy")
+BURGUNDY = gr.themes.Color(c50="#fdf2f4", c100="#fbe5e9", c200="#f5c6d0", c300="#ec9aab", c400="#d9637c",
+                           c500="#9b2743", c600="#801f38", c700="#661a2e", c800="#4f1524", c900="#3d1019",
+                           c950="#2e0a14", name="burgundy")
 
-   demo.launch(theme=gr.themes.Base(primary_hue=BURGUNDY, neutral_hue="slate"), css=CSS,
-               js="() => document.body.classList.add('dark')")
+demo.launch(theme=gr.themes.Base(primary_hue=BURGUNDY, neutral_hue="slate"), css=CSS,
+            js="() => document.body.classList.add('dark')")