training_config.json

{
  "model_name": "meta-llama/Llama-3.2-3B-Instruct",
  "dataset_name": "ShivomH/MentalHealth-Support",
  "output_dir": "./Elixir-MentalHealth-Llama3.2",
  "adapter_dir": "./Elixir-MentalHealth-LoRA",
  "merged_model_dir": "./Elixir-MentalHealth-Llama3.2-v2",
  "cache_dir": "./cache",
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "target_modules": [
    "q_proj",
    "v_proj",
    "k_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "num_train_epochs": 4,
  "per_device_train_batch_size": 8,
  "per_device_eval_batch_size": 8,
  "gradient_accumulation_steps": 4,
  "learning_rate": 3e-05,
  "warmup_ratio": 0.05,
  "max_seq_length": 1024,
  "optim": "paged_adamw_32bit",
  "lr_scheduler_type": "cosine",
  "gradient_checkpointing": true,
  "bf16": true,
  "tf32": true,
  "logging_steps": 50,
  "save_steps": 400,
  "eval_steps": 200,
  "save_total_limit": 3,
  "use_wandb": true,
  "wandb_project": "Elixir-v2-mentalhealth-finetune",
  "wandb_run_name": "llama3.2-lora-20250925_174907",
  "push_to_hub": false,
  "hub_model_id": "ShivomH/Elixir-Health-v2",
  "hub_private": true
}