Upload model artifacts and README

README.md

@@ -1,91 +1,65 @@
 ---
-license: mit
+library_name: transformers
+pipeline_tag: text-classification
 tags:
 - text-classification
-- answering-machine-detection
-- bert-tiny
-- binary-classification
-- call-center
-- voice-processing
-pipeline_tag: text-classification
+- voicemail-detection
+- bert
+- pytorch
+license: apache-2.0
 ---
 
-# BERT-Tiny AMD Classifier
-
-A lightweight BERT-Tiny model fine-tuned for Answering Machine Detection (AMD) in call center environments.
-
-## Model Description
+# Voicemail Detection Model (3-Utterance)
 
-This model is based on `prajjwal1/bert-tiny` and fine-tuned to classify phone call transcripts as either human or machine (answering machine/voicemail) responses. It's designed for real-time call center applications where quick and accurate detection of answering machines is crucial.
-
-## Model Architecture
-
-- **Base Model**: `prajjwal1/bert-tiny` (2 layers, 128 hidden size, 2 attention heads)
-- **Total Parameters**: ~4.4M (lightweight and efficient)
-- **Input**: User transcript text (max 128 tokens)
-- **Output**: Single logit with sigmoid activation for binary classification
-- **Loss Function**: BCEWithLogitsLoss with positive weight for class imbalance
+Binary classification model to detect voicemail vs human on phone calls.
 
 ## Performance
 
-- **Validation Accuracy**: 93.94%
-- **Precision**: 92.75%
-- **Recall**: 87.27%
-- **F1-Score**: 89.93%
-- **Training Device**: MPS (Apple Silicon GPU)
-- **Best Epoch**: 15 (with early stopping)
+### Validation Set
+- Accuracy: 0.9703
+- Precision: 0.9005
+- Recall: 0.9794
+- F1: 0.9383
 
-## Training Data
+### Test Set
+- Accuracy: 0.8353
+- Precision: 0.6678
+- Recall: 0.9895
+- F1: 0.7975
 
-- **Total Samples**: 3,548 phone call transcripts
-- **Training Set**: 2,838 samples
-- **Validation Set**: 710 samples
-- **Class Distribution**: 30.8% machine calls, 69.2% human calls
-- **Source**: ElevateNow call center data
+## Details
 
-## Usage
+Base: prajjwal1/bert-tiny
+Threshold: 0.1153
+Training: 2025-10-04
 
-### Basic Inference
+## Usage
 
 ```python
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import torch
 
-# Load model and tokenizer
-model = AutoModelForSequenceClassification.from_pretrained("Adya662/bert-tiny-amd")
-tokenizer = AutoTokenizer.from_pretrained("Adya662/bert-tiny-amd")
+model_id = "Adya662/bert-tiny-amd"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForSequenceClassification.from_pretrained(model_id)
+model.eval()
 
-# Prepare input
-text = "Hello, this is John speaking"
-inputs = tokenizer(text, return_tensors="pt", max_length=128, truncation=True, padding=True)
+text = "Hi you've reached voicemail"
+encoding = tokenizer(
+    text,
+    return_tensors='pt',
+    max_length=128,
+    padding='max_length',
+    truncation=True
+)
 
-# Make prediction
 with torch.no_grad():
-    outputs = model(**inputs)
-    logits = outputs.logits.squeeze(-1)
-    probability = torch.sigmoid(logits).item()
-    is_machine = probability >= 0.5
-
-print(f"Prediction: {'Machine' if is_machine else 'Human'}")
-print(f"Confidence: {probability:.4f}")
+    outputs = model(**encoding)
+    # Assuming label 1 = voicemail (update if different)
+    probs = torch.softmax(outputs.logits, dim=-1)
+    probability = probs[0, 1].item()
+
+optimal_threshold = 0.1153
+prediction = "voicemail" if probability >= optimal_threshold else "human"
+print({"probability": probability, "prediction": prediction})
 ```
-
-## Training Details
-
-- **Optimizer**: AdamW with weight decay (0.01)
-- **Learning Rate**: 3e-5 with linear scheduling
-- **Batch Size**: 32
-- **Epochs**: 15 (with early stopping)
-- **Early Stopping**: Patience of 3 epochs
-- **Class Imbalance**: Handled with positive weight
-
-## Limitations
-
-- Trained on English phone call transcripts
-- May not generalize well to other languages or domains
-- Performance may vary with different transcription quality
-- Designed for short utterances (max 128 tokens)
-
-## License
-
-MIT License - see LICENSE file for details.

model_metadata.json

@@ -0,0 +1,30 @@
+{
+  "model_name": "prajjwal1/bert-tiny",
+  "max_length": 128,
+  "optimal_threshold": 0.11530892550945282,
+  "val_metrics": {
+    "accuracy": 0.9702797202797203,
+    "precision": 0.9004739336492891,
+    "recall": 0.979381443298969,
+    "f1": 0.9382716049382717
+  },
+  "test_metrics": {
+    "accuracy": 0.8353344768439108,
+    "precision": 0.6678445229681979,
+    "recall": 0.9895287958115183,
+    "f1": 0.7974683544303798,
+    "confusion_matrix": [
+      [
+        298,
+        94
+      ],
+      [
+        2,
+        189
+      ]
+    ]
+  },
+  "dropout_rate": 0.2,
+  "training_date": "2025-10-04 02:43:10",
+  "hidden_size": 128
+}

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5f8c3c949e8963d27748803fe785af04652da64704533cfcdcdeae7505f0d328
-size 17598379
+oid sha256:ad78e66f593f90cb4c185ca79d87eff358815a94e8fbc373278971a9ae9eec37
+size 18493158

tokenizer.json

@@ -1,7 +1,21 @@
 {
   "version": "1.0",
-  "truncation": null,
-  "padding": null,
+  "truncation": {
+    "direction": "Right",
+    "max_length": 128,
+    "strategy": "LongestFirst",
+    "stride": 0
+  },
+  "padding": {
+    "strategy": {
+      "Fixed": 128
+    },
+    "direction": "Right",
+    "pad_to_multiple_of": null,
+    "pad_id": 0,
+    "pad_type_id": 0,
+    "pad_token": "[PAD]"
+  },
   "added_tokens": [
     {
       "id": 0,