neginashz/star-sft-intellect-instruct-5

base_model: PrimeIntellect/INTELLECT-1-Instruct
trust_remote_code: true
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

gpu_memory_limit: 

deepspeed: deepspeed_configs/zero2.json

load_in_8bit: 
load_in_4bit:
strict: false

chat_template: llama3
datasets:
  - path: neginashz/rationale-llama-chat-dataset
    type: chat_template
    chat_template: llama3
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      system:
        - system
      user:
        - user
      assistant:
        - assistant
    #roles_to_train: ["assistant"]  # default
    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
    # - all: train on all EOS tokens
    # - turn (default): train on the EOS token at the end of each trainable turn
    # - last: train on the last EOS token in the conversation
    #train_on_eos: turn

    
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./star-sft-intellect-5

sequence_len: 4096
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true


wandb_project: star-sft-intellect-instruct-5
wandb_entity: 
wandb_watch:
wandb_name: 
wandb_log_model: 

gradient_checkpointing: true
#gradient_clipping: true
gradient_accumulation_steps: 1
#batch_size: 1
micro_batch_size: 1

num_epochs: 1

optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00002

train_on_inputs: false
group_by_length: false

bf16: true
fp16: false
tf32: false

logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps:
eval_steps: 
save_steps:

evals_per_epoch: 16
saves_per_epoch: 4
eval_max_new_tokens: 128

debug:

weight_decay:
fsdp:
fsdp_config:

hub_model_id: neginashz/star-sft-intellect-instruct-5
hub_strategy: 
early_stopping_patience:

resume_from_checkpoint:
auto_resume_from_checkpoints: true

#special_tokens:
#   pad_token: <|end_of_text|>