base_model: PrimeIntellect/INTELLECT-1-Instruct
trust_remote_code: true
model_type: AutoModelForCausalLM
tokenizer_config: meta-llama/Llama-3.1-8B-Instruct

#model_type: LlamaForCausalLM
#tokenizer_type: llama3

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-6

sequence_len: 8192
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true


wandb_project: star-sft-intellect-instruct-6
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: 8
saves_per_epoch: 2
eval_max_new_tokens: 128

debug:

weight_decay:
fsdp:
fsdp_config:

hub_model_id: neginashz/star-sft-intellect-instruct-6
hub_strategy: 
early_stopping_patience:

resume_from_checkpoint:
auto_resume_from_checkpoints: false

special_tokens:
   pad_token: <|finetune_right_pad_id|>
   eos_token": <|eot_id|>

star-sft-intellect-instruct-6

This model is a fine-tuned version of PrimeIntellect/INTELLECT-1-Instruct on the neginashz/rationale-llama-chat-dataset dataset. It achieves the following results on the evaluation set:

Loss: 0.3380

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 4
total_train_batch_size: 4
total_eval_batch_size: 4
optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 3
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.4428	0.1261	14	0.4024
0.433	0.2523	28	0.3939
0.4197	0.3784	42	0.3799
0.4083	0.5045	56	0.3679
0.357	0.6306	70	0.3534
0.3623	0.7568	84	0.3435
0.3645	0.8829	98	0.3380

Framework versions

Transformers 4.47.1
Pytorch 2.5.1+cu124
Datasets 3.1.0
Tokenizers 0.21.0