RCDS1.yaml

base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Guytron/RosettaCodeDataSet1
    type: json  # Assuming the dataset is in JSON format
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./qlora-out-rosetta

adapter: qlora
lora_model_dir:

sequence_len: 2048  # Increased to accommodate potentially longer code samples
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: rosetta-code-training
wandb_entity:
wandb_watch:
wandb_name: rosetta-code-run-1
wandb_log_model:

mlflow_experiment_name: rosetta-code-experiment

gradient_accumulation_steps: 4  # Increased to handle larger dataset
micro_batch_size: 2  # Adjusted based on your GPU memory
num_epochs: 3
max_steps: -1  # Set to -1 to train on the entire dataset
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: true  # Changed to true for efficiency with varying length samples
bf16: false
fp16: true
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 10
xformers_attention:
flash_attention: false

warmup_steps: 100  # Increased for a larger dataset
evals_per_epoch: 1
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.01  # Added some weight decay for regularization
fsdp:
fsdp_config:
special_tokens: