architecture:
    backbone_dtype: bfloat16
    force_embedding_gradients: false
    gradient_checkpointing: true
    intermediate_dropout: 0.0
    pretrained: true
    pretrained_weights: ''
    siamese: true
augmentation:
    neftune_noise_alpha: 0.0
    random_parent_probability: 0.0
    skip_parent_probability: 0.0
    token_mask_probability: 0.0
dataset:
    add_eos_token_to_answer: false
    add_eos_token_to_prompt: false
    add_eos_token_to_system: false
    answer_column: label
    chatbot_author: H2O.ai
    chatbot_name: h2oGPT
    data_sample: 1.0
    data_sample_choice: []
    limit_chained_samples: false
    mask_prompt_labels: true
    num_classes: 3
    parent_id_column: None
    personalize: false
    prompt_column:
    - text
    system_column: None
    text_answer_separator: ''
    text_prompt_start: ''
    text_system_start: ''
    train_dataframe: /home/philipp/h2o-llmstudio/data/user/lmsys/train_siamese_v1.pq
    validation_dataframe: /home/philipp/h2o-llmstudio/data/user/lmsys/val_siamese_v1.pq
    validation_size: 0.1
    validation_strategy: custom
environment:
    compile_model: false
    deepspeed_allgather_bucket_size: 1000000
    deepspeed_method: ZeRO2
    deepspeed_reduce_bucket_size: 1000000
    deepspeed_stage3_param_persistence_threshold: 1000000
    deepspeed_stage3_prefetch_bucket_size: 1000000
    find_unused_parameters: false
    gpus:
    - '0'
    - '1'
    - '2'
    huggingface_branch: main
    mixed_precision: true
    mixed_precision_dtype: bfloat16
    number_of_workers: 8
    seed: -1
    trust_remote_code: true
    use_deepspeed: false
experiment_name: important-monkey-siamese-v2
llm_backbone: microsoft/Phi-3-mini-4k-instruct
logging:
    logger: Neptune
    neptune_project: Zoo/kaggle-lmsys
output_directory: /home/philipp/h2o-llmstudio/output/user/important-monkey-siamese-v2/
prediction:
    batch_size_inference: 0
    metric: LogLoss
problem_type: text_causal_classification_modeling
tokenizer:
    add_prompt_answer_tokens: false
    max_length: 4096
    max_length_answer: 4096
    max_length_prompt: 4096
    padding_quantile: 1.0
    use_fast: true
training:
    batch_size: 2
    differential_learning_rate: 1.0e-05
    differential_learning_rate_layers:
    - classification_head
    drop_last_batch: true
    epochs: 1
    evaluate_before_training: false
    evaluation_epochs: 0.25
    grad_accumulation: 1
    gradient_clip: 0.0
    learning_rate: 0.0001
    lora: true
    lora_alpha: 32
    lora_dropout: 0.0
    lora_r: 16
    lora_target_modules: ''
    loss_function: CrossEntropyLoss
    optimizer: AdamW
    save_checkpoint: last
    schedule: Cosine
    train_validation_data: false
    use_flash_attention_2: true
    warmup_epochs: 0.0
    weight_decay: 0.0