Upload JengaAI trained multi-task NLP model

checkpoints/best/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 2,
+  "metrics": {
+    "fraud_detection_fraud_accuracy": 0.5,
+    "fraud_detection_fraud_precision": 0.8333333333333334,
+    "fraud_detection_fraud_recall": 0.5,
+    "fraud_detection_fraud_f1": 0.5,
+    "sentiment_analysis_sentiment_accuracy": 0.6666666666666666,
+    "sentiment_analysis_sentiment_precision": 0.8333333333333334,
+    "sentiment_analysis_sentiment_recall": 0.6666666666666666,
+    "sentiment_analysis_sentiment_f1": 0.6666666666666666,
+    "call_quality_opening_precision": 0.966,
+    "call_quality_opening_recall": 0.966,
+    "call_quality_opening_f1": 0.966,
+    "call_quality_listening_precision": 0.8960259935016246,
+    "call_quality_listening_recall": 0.9421813403416557,
+    "call_quality_listening_f1": 0.9185242121445042,
+    "call_quality_proactiveness_precision": 0.7589473684210526,
+    "call_quality_proactiveness_recall": 0.828735632183908,
+    "call_quality_proactiveness_f1": 0.7923076923076923,
+    "call_quality_resolution_precision": 0.9283206553540082,
+    "call_quality_resolution_recall": 0.8880492583263364,
+    "call_quality_resolution_f1": 0.9077385209555142,
+    "call_quality_hold_precision": 0.7696629213483146,
+    "call_quality_hold_recall": 0.6650485436893204,
+    "call_quality_hold_f1": 0.7135416666666666,
+    "call_quality_closing_precision": 0.868,
+    "call_quality_closing_recall": 0.868,
+    "call_quality_closing_f1": 0.868,
+    "eval_loss": 1.9685399062990203,
+    "train_loss_avg": 1.9783029844836584
+  },
+  "is_best": true
+}

checkpoints/best/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8b696a7fa15b75f2ac27d2c8f26442591698f223e5dce2e38de5190ba7db982
+size 503468023

checkpoints/epoch_6/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 6,
+  "metrics": {
+    "fraud_detection_fraud_accuracy": 1.0,
+    "fraud_detection_fraud_precision": 1.0,
+    "fraud_detection_fraud_recall": 1.0,
+    "fraud_detection_fraud_f1": 1.0,
+    "sentiment_analysis_sentiment_accuracy": 0.6666666666666666,
+    "sentiment_analysis_sentiment_precision": 0.8333333333333334,
+    "sentiment_analysis_sentiment_recall": 0.6666666666666666,
+    "sentiment_analysis_sentiment_f1": 0.6666666666666666,
+    "call_quality_opening_precision": 0.955,
+    "call_quality_opening_recall": 0.955,
+    "call_quality_opening_f1": 0.955,
+    "call_quality_listening_precision": 0.8877526174823472,
+    "call_quality_listening_recall": 0.9582128777923784,
+    "call_quality_listening_f1": 0.9216380182002022,
+    "call_quality_proactiveness_precision": 0.7404233870967742,
+    "call_quality_proactiveness_recall": 0.8442528735632184,
+    "call_quality_proactiveness_f1": 0.7889366272824919,
+    "call_quality_resolution_precision": 0.8931277056277056,
+    "call_quality_resolution_recall": 0.9238734956619088,
+    "call_quality_resolution_f1": 0.9082404732425368,
+    "call_quality_hold_precision": 0.7391304347826086,
+    "call_quality_hold_recall": 0.6601941747572816,
+    "call_quality_hold_f1": 0.6974358974358974,
+    "call_quality_closing_precision": 0.857,
+    "call_quality_closing_recall": 0.857,
+    "call_quality_closing_f1": 0.857,
+    "eval_loss": 2.2391653188511613,
+    "train_loss_avg": 1.4627334099144689
+  },
+  "is_best": false
+}

checkpoints/epoch_6/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d69f2d1a5baa2f1f073d43ca46218f420a9f73473bd6375ecd45053b2001b834
+size 503468023

checkpoints/epoch_7/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 7,
+  "metrics": {
+    "fraud_detection_fraud_accuracy": 1.0,
+    "fraud_detection_fraud_precision": 1.0,
+    "fraud_detection_fraud_recall": 1.0,
+    "fraud_detection_fraud_f1": 1.0,
+    "sentiment_analysis_sentiment_accuracy": 0.3333333333333333,
+    "sentiment_analysis_sentiment_precision": 0.1111111111111111,
+    "sentiment_analysis_sentiment_recall": 0.3333333333333333,
+    "sentiment_analysis_sentiment_f1": 0.16666666666666666,
+    "call_quality_opening_precision": 0.955,
+    "call_quality_opening_recall": 0.955,
+    "call_quality_opening_f1": 0.955,
+    "call_quality_listening_precision": 0.8899975663178389,
+    "call_quality_listening_recall": 0.9611038107752956,
+    "call_quality_listening_f1": 0.9241849886277483,
+    "call_quality_proactiveness_precision": 0.7540134645261523,
+    "call_quality_proactiveness_recall": 0.8367816091954023,
+    "call_quality_proactiveness_f1": 0.7932443475892127,
+    "call_quality_resolution_precision": 0.9047486809219661,
+    "call_quality_resolution_recall": 0.9118387909319899,
+    "call_quality_resolution_f1": 0.9082798996375802,
+    "call_quality_hold_precision": 0.7282051282051282,
+    "call_quality_hold_recall": 0.6893203883495146,
+    "call_quality_hold_f1": 0.7082294264339152,
+    "call_quality_closing_precision": 0.853,
+    "call_quality_closing_recall": 0.853,
+    "call_quality_closing_f1": 0.853,
+    "eval_loss": 2.245173363469717,
+    "train_loss_avg": 1.345002403695299
+  },
+  "is_best": false
+}

checkpoints/epoch_7/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a65dcb9159567e935bd33163026780135a1e88e77f39e6141193cc50a2aacc3
+size 503468023

encoder_config/config.json

@@ -0,0 +1,27 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "RobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "is_decoder": false,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.0.0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

experiment_config.yaml

@@ -0,0 +1,131 @@
+project_name: security_nlp_3task
+tasks:
+- name: fraud_detection
+  type: single_label_classification
+  data_path: fixtures/sample_classification.jsonl
+  heads:
+  - name: fraud
+    num_labels: 2
+    weight: 1.0
+    dropout: 0.1
+  text_column: text
+  label_column: label
+  label_maps: null
+- name: sentiment_analysis
+  type: single_label_classification
+  data_path: fixtures/sample_sentiment.jsonl
+  heads:
+  - name: sentiment
+    num_labels: 3
+    weight: 1.0
+    dropout: 0.1
+  text_column: text
+  label_column: label
+  label_maps: null
+- name: call_quality
+  type: question_answering
+  data_path: /content/JengaAI/fixtures/synthetic_qa_metrics_data_v01x.json
+  heads:
+  - name: opening
+    num_labels: 1
+    weight: 1.0
+    dropout: 0.1
+  - name: listening
+    num_labels: 5
+    weight: 1.5
+    dropout: 0.1
+  - name: proactiveness
+    num_labels: 3
+    weight: 1.0
+    dropout: 0.1
+  - name: resolution
+    num_labels: 5
+    weight: 2.0
+    dropout: 0.1
+  - name: hold
+    num_labels: 2
+    weight: 0.5
+    dropout: 0.1
+  - name: closing
+    num_labels: 1
+    weight: 1.0
+    dropout: 0.1
+  text_column: transcript
+  label_column: labels
+  label_maps:
+    opening:
+      0: greeting
+    listening:
+      0: acknowledgment
+      1: empathy
+      2: clarification
+      3: active_listening
+      4: patience
+    proactiveness:
+      0: initiative
+      1: follow_up
+      2: suggestions
+    resolution:
+      0: identified_issue
+      1: provided_solution
+      2: confirmed_resolution
+      3: set_expectations
+      4: offered_alternatives
+    hold:
+      0: asked_permission
+      1: explained_reason
+    closing:
+      0: proper_farewell
+model:
+  base_model: roberta-base
+  hidden_size: 768
+  dropout: 0.1
+  fusion:
+    type: attention
+    dropout: 0.1
+    use_residual: true
+    num_attention_heads: 1
+    gate_init_value: 0.5
+  freeze_encoder_layers: 0
+  gradient_checkpointing: false
+tokenizer:
+  max_length: 256
+  padding: max_length
+  truncation: true
+training:
+  output_dir: ./results/roberta-base-3task_fusion
+  learning_rate: 2.0e-05
+  batch_size: 8
+  eval_batch_size: 8
+  num_epochs: 15
+  weight_decay: 0.01
+  warmup_steps: 20
+  max_grad_norm: 1.0
+  gradient_accumulation_steps: 1
+  use_amp: false
+  device: auto
+  task_sampling: proportional
+  temperature: 2.0
+  early_stopping_patience: 5
+  metric_for_best_model: eval_loss
+  greater_is_better: false
+  logging:
+    service: tensorboard
+    experiment_name: security_3task_fusion
+    tracking_uri: null
+    log_every_n_steps: 5
+  checkpoint:
+    save_every_n_epochs: 1
+    save_best: true
+    max_checkpoints: 2
+  data:
+    test_size: 0.2
+    seed: 42
+    num_workers: 0
+    pin_memory: true
+    pii_redaction:
+      enabled: false
+      strategy: mask
+      detect_types: null
+      hash_salt: jenga-ai-pii
+      log_detections: true

logs/security_3task_fusion/events.out.tfevents.1771237843.71520f106676.2965.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:767b24c6c27e9850bc86cf0eb369b6e7f678ab706012980d26fe9db54bcc9bc0
+size 55381

metadata.json

@@ -0,0 +1,16 @@
+{
+  "base_model": "roberta-base",
+  "hidden_size": 768,
+  "num_tasks": 3,
+  "task_names": [
+    "fraud_detection",
+    "sentiment_analysis",
+    "call_quality"
+  ],
+  "task_types": [
+    "single_label_classification",
+    "single_label_classification",
+    "question_answering"
+  ],
+  "has_fusion": true
+}

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a65dcb9159567e935bd33163026780135a1e88e77f39e6141193cc50a2aacc3
+size 503468023