Upload JengaAI trained multi-task NLP model

checkpoints/best/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 4,
+  "metrics": {
+    "fraud_detection_fraud_accuracy": 0.75,
+    "fraud_detection_fraud_precision": 0.875,
+    "fraud_detection_fraud_recall": 0.75,
+    "fraud_detection_fraud_f1": 0.7666666666666667,
+    "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": 0.963,
+    "call_quality_opening_recall": 0.963,
+    "call_quality_opening_f1": 0.963,
+    "call_quality_listening_precision": 0.8913311928871326,
+    "call_quality_listening_recall": 0.9484888304862024,
+    "call_quality_listening_f1": 0.919022154316272,
+    "call_quality_proactiveness_precision": 0.7491271820448878,
+    "call_quality_proactiveness_recall": 0.8632183908045977,
+    "call_quality_proactiveness_f1": 0.8021361815754339,
+    "call_quality_resolution_precision": 0.9153363740022805,
+    "call_quality_resolution_recall": 0.8986845787853345,
+    "call_quality_resolution_f1": 0.9069340488631549,
+    "call_quality_hold_precision": 0.7804878048780488,
+    "call_quality_hold_recall": 0.46601941747572817,
+    "call_quality_hold_f1": 0.5835866261398176,
+    "call_quality_closing_precision": 0.863,
+    "call_quality_closing_recall": 0.863,
+    "call_quality_closing_f1": 0.863,
+    "eval_loss": 1.9891562443513138,
+    "train_loss_avg": 1.9425991174011004
+  },
+  "is_best": true
+}

checkpoints/best/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9f04c79105542cda3168410c549cfca104c566e5dd04c1c66496cfb38ccc0bd
+size 450684151

checkpoints/epoch_8/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 8,
+  "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": 0.961,
+    "call_quality_opening_recall": 0.961,
+    "call_quality_opening_f1": 0.961,
+    "call_quality_listening_precision": 0.8937638649248213,
+    "call_quality_listening_recall": 0.9529566360052563,
+    "call_quality_listening_f1": 0.9224116001017553,
+    "call_quality_proactiveness_precision": 0.7430112800392349,
+    "call_quality_proactiveness_recall": 0.8706896551724138,
+    "call_quality_proactiveness_f1": 0.8017994178354062,
+    "call_quality_resolution_precision": 0.9060665362035225,
+    "call_quality_resolution_recall": 0.9070808844108592,
+    "call_quality_resolution_f1": 0.9065734265734265,
+    "call_quality_hold_precision": 0.75177304964539,
+    "call_quality_hold_recall": 0.5145631067961165,
+    "call_quality_hold_f1": 0.6109510086455331,
+    "call_quality_closing_precision": 0.851,
+    "call_quality_closing_recall": 0.851,
+    "call_quality_closing_f1": 0.851,
+    "eval_loss": 2.139722445148688,
+    "train_loss_avg": 1.5664203519385957
+  },
+  "is_best": false
+}

checkpoints/epoch_8/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53cd910936c79d638b139c62e284889edfb889b87ff170d7289fd91085855df4
+size 450684151

checkpoints/epoch_9/checkpoint_meta.json

@@ -0,0 +1,34 @@
+{
+  "epoch": 9,
+  "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": 0.963,
+    "call_quality_opening_recall": 0.963,
+    "call_quality_opening_f1": 0.963,
+    "call_quality_listening_precision": 0.8982245561390347,
+    "call_quality_listening_recall": 0.9440210249671485,
+    "call_quality_listening_f1": 0.920553562275756,
+    "call_quality_proactiveness_precision": 0.7580813347236705,
+    "call_quality_proactiveness_recall": 0.835632183908046,
+    "call_quality_proactiveness_f1": 0.7949699289229087,
+    "call_quality_resolution_precision": 0.9113707609005415,
+    "call_quality_resolution_recall": 0.8950461796809404,
+    "call_quality_resolution_f1": 0.9031347077096865,
+    "call_quality_hold_precision": 0.7575757575757576,
+    "call_quality_hold_recall": 0.4854368932038835,
+    "call_quality_hold_f1": 0.591715976331361,
+    "call_quality_closing_precision": 0.848,
+    "call_quality_closing_recall": 0.848,
+    "call_quality_closing_f1": 0.848,
+    "eval_loss": 2.179060117785747,
+    "train_loss_avg": 1.5035246791362407
+  },
+  "is_best": false
+}

checkpoints/epoch_9/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:debc0005634ff2b1e3728ced2194492c727ae1dfd1de9a52a8ae313c5fcf3b85
+size 450684151

encoder_config/config.json

@@ -0,0 +1,30 @@
+{
+  "add_cross_attention": false,
+  "architectures": [
+    "XLMRobertaForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "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": "xlm-roberta",
+  "num_attention_heads": 6,
+  "num_hidden_layers": 8,
+  "output_past": true,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.0.0",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 70006
+}

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: 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: castorini/afriberta_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/3task_fusion
+  learning_rate: 2.0e-05
+  batch_size: 16
+  eval_batch_size: 16
+  num_epochs: 12
+  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.1771217131.7207cada2386.3311.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bf8c2b49a191f57676c7419cc02f6bc4f91c9f07c53762ed936e5bcf502144e
+size 44553

metadata.json

@@ -0,0 +1,16 @@
+{
+  "base_model": "castorini/afriberta_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:debc0005634ff2b1e3728ced2194492c727ae1dfd1de9a52a8ae313c5fcf3b85
+size 450684151