### LLaMA-Factory SFT Training Config - QLoRA (Ultra Minimal Hardware) ###
### For session-level reflection training ###
### Can run on single GPU with ~12GB VRAM ###

### Model
model_name_or_path: /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct

### Method - QLoRA (4-bit quantization + LoRA)
stage: sft
do_train: true
finetuning_type: lora
quantization_bit: 4
quantization_method: bitsandbytes

### LoRA Config
lora_rank: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target: all

### Dataset
dataset: sft_reflection
dataset_dir: /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/training
template: llama3
cutoff_len: 4096

### Output
output_dir: /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/training/outputs/sft_reflection_qlora

### Training - Single GPU friendly
per_device_train_batch_size: 1
gradient_accumulation_steps: 64
learning_rate: 2.0e-5
num_train_epochs: 4.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true

### Logging
logging_steps: 10
save_steps: 100
save_total_limit: 3