training_args.json

{
  "model": "langtech-innovation/7b-tools-v3",
  "data": "data/training/impulse_training.jsonl",
  "output_dir": "models/impulse-7b-tools-v3-ft",
  "epochs": 5,
  "batch_size": 1,
  "gradient_accumulation_steps": 16,
  "learning_rate": 0.0001,
  "max_length": 4096,
  "quantize": "4bit",
  "warmup_steps": 100,
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "eval_split": 0.1,
  "seed": 42,
  "use_wandb": false,
  "early_stopping_patience": 3,
  "eval_accumulation_steps": 4,
  "skip_perplexity": true
}