first commit

inspect_preds.py

@@ -0,0 +1,46 @@
+import pandas as pd
+import torch
+from transformers import ByT5Tokenizer, T5ForConditionalGeneration
+
+# 1. Load your fine-tuned model and tokenizer
+model_path = "./byt5-taglish-nli-final-v3"
+print(f"Loading model from {model_path}...")
+
+# local_files_only=True bypasses the cluster's DNS issues
+tokenizer = ByT5Tokenizer.from_pretrained(model_path, local_files_only=True)
+model = T5ForConditionalGeneration.from_pretrained(model_path, local_files_only=True)
+
+# 2. Load the first 10 samples from the benchmark dataset
+print("Loading benchmark dataset...")
+df_test = pd.read_csv("benchmark_dataset.csv")
+sample_df = df_test.head(50)
+
+# 3. Generate and decode predictions
+print("\n" + "="*50)
+print("RAW MODEL PREDICTIONS")
+print("="*50)
+
+label_map = {0: "entailment", 1: "neutral", 2: "contradiction"}
+
+for index, row in sample_df.iterrows():
+    # Recreate the exact prompt format used during training
+    # input_text = f"nli premise: {row['s1']} hypothesis: {row['s2']}"
+    input_text = f"Context: {row['s1']} Statement: {row['s2']} Question: Does the context entail, contradict, or remain neutral to the statement? Answer:"
+
+    # Tokenize the input
+    inputs = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True)
+    
+    # Generate the prediction bytes
+    # max_new_tokens=25 is added to fix the warning you saw earlier. 
+    # "contradiction" takes roughly 13 byte-tokens in ByT5.
+    outputs = model.generate(**inputs, max_new_tokens=25)
+    
+    # Decode the raw bytes back into readable text
+    raw_prediction = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    
+    true_label = label_map.get(row['label'], "unknown")
+
+    print(f"\nSample {index + 1}:")
+    print(f"Input:      {input_text}")
+    print(f"Expected:   '{true_label}'")
+    print(f"Model Says: '{raw_prediction}'")