初始化项目，由ModelHub XC社区提供模型

Model: FogTeams/experiment-105-model-consolidation-itr-1 Source: Original Platform
2026-06-03 20:44:18 +08:00
commit d76e10df98
12 changed files with 2744 additions and 0 deletions
--- a/cfg.yaml
+++ b/cfg.yaml
@@ -0,0 +1,116 @@
+architecture:
+    backbone_dtype: bfloat16
+    gradient_checkpointing: true
+    intermediate_dropout: 0.0
+    pretrained: true
+    pretrained_weights: ''
+augmentation:
+    neftune_noise_alpha: 5.0
+    random_parent_probability: 0.0
+    skip_parent_probability: 0.0
+    token_mask_probability: 0.0
+dataset:
+    add_eos_token_to_answer: true
+    add_eos_token_to_prompt: true
+    add_eos_token_to_system: true
+    answer_column: ground_truth
+    chatbot_author: H2O.ai
+    chatbot_name: h2oGPT
+    data_sample: 0.5
+    data_sample_choice:
+    - Train
+    id_column: None
+    limit_chained_samples: false
+    mask_prompt_labels: true
+    only_last_answer: false
+    parent_id_column: None
+    personalize: false
+    prompt_column:
+    - prompt
+    prompt_column_separator: \\n\\n
+    system_column: None
+    text_answer_separator: <|answer|>
+    text_prompt_start: <|prompt|>
+    text_system_start: <|system|>
+    train_dataframe: /home/ubuntu/h2o-llmstudio/data/user/data_for_LLM_Studio_exp_105_model_consolidation_itr_1.2/train_data_for_LLM_Studio_exp_105_model_consolidation_itr_1.csv
+    validation_dataframe: /home/ubuntu/h2o-llmstudio/data/user/data_for_LLM_Studio_exp_105_model_consolidation_itr_1.2/validation_data_for_LLM_Studio_exp_105_model_consolidation_itr_1.csv
+    validation_size: 0.01
+    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'
+    huggingface_branch: main
+    mixed_precision: false
+    mixed_precision_dtype: bfloat16
+    number_of_workers: 8
+    seed: 77
+    trust_remote_code: true
+    use_deepspeed: false
+experiment_name: experiment_105_model_consolidation_itr_1
+llm_backbone: meta-llama/Llama-3.2-3B
+logging:
+    log_all_ranks: false
+    log_step_size: absolute
+    logger: None
+    neptune_project: ''
+    wandb_entity: ''
+    wandb_project: ''
+output_directory: /home/ubuntu/h2o-llmstudio/output/user/experiment_105_model_consolidation_itr_1/
+prediction:
+    batch_size_inference: 4
+    do_sample: true
+    max_length_inference: 201
+    max_time: 0.0
+    metric: BLEU
+    metric_gpt_model: gpt-3.5-turbo-0301
+    metric_gpt_template: general
+    min_length_inference: 1
+    num_beams: 1
+    num_history: 4
+    repetition_penalty: 1.2
+    stop_tokens: ''
+    temperature: 0.3
+    top_k: 50
+    top_p: 0.95
+problem_type: text_causal_language_modeling
+tokenizer:
+    add_prompt_answer_tokens: false
+    max_length: 4000
+    padding_quantile: 1.0
+    tokenizer_kwargs: '{"use_fast": true, "add_prefix_space": false}'
+training:
+    attention_implementation: flash_attention_2
+    batch_size: 6
+    differential_learning_rate: 1.0e-05
+    differential_learning_rate_layers: []
+    drop_last_batch: true
+    epochs: 2
+    evaluate_before_training: false
+    evaluation_epochs: 1.0
+    freeze_layers: []
+    grad_accumulation: 2
+    gradient_clip: 0.0
+    learning_rate: 0.0001
+    lora: true
+    lora_alpha: 32
+    lora_dropout: 0.05
+    lora_r: 8
+    lora_target_modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
+    lora_unfreeze_layers: []
+    loss_function: TokenAveragedCrossEntropy
+    min_learning_rate_ratio: 0.0
+    optimizer: AdamW8bit
+    save_checkpoint: best
+    schedule: Cosine
+    train_validation_data: false
+    use_dora: true
+    use_rslora: false
+    warmup_epochs: 0.05
+    weight_decay: 0.01