model_name_or_path: ali-elganzory/1.7b-Comma0.1-300BT-longsft_16k
dtype: bfloat16
attn_implementation: flash_attention_2
trust_remote_code: true
chat_template: "{%- for message in messages -%}\n\t{%- if message[\"role\"] == \"\
  system\" -%}\n\t\t{{- \"<|system|>\\n\" + message[\"content\"] + \"\\n\" -}}\n\t\
  {%- elif message[\"role\"] == \"user\" -%}\n\t\t{{- \"<|user|>\\n\" + message[\"\
  content\"] + \"\\n\" -}}\n\t{%- elif message[\"role\"] == \"assistant\" -%}\n\t\t\
  {%- if not loop.last -%}\n\t\t\t{{- \"<|assistant|>\\n\" -}}{% generation %}{{-\
  \ message[\"content\"] + eos_token -}}{% endgeneration %}{{- \"\\n\" -}}\n\t\t{%-\
  \ else -%}\n\t\t\t{{- \"<|assistant|>\\n\" -}}{% generation %}{{- message[\"content\"\
  ] + eos_token -}}{% endgeneration %}\n\t\t{%- endif -%}\n\t{%- endif -%}\n\t{%-\
  \ if loop.last and add_generation_prompt -%}\n\t\t{{- \"<|assistant|>\\n\" -}}\n\
  \t{%- endif -%}\n{%- endfor -%}\n"
additional_special_tokens:
- <|system|>
- <|user|>
- <|assistant|>
dataset_mixture:
  datasets:
  - id: ali-elganzory/tulu-3-sft-mixture-decontaminated
    config: default
    split: train
    columns:
    - messages
dataset_num_proc: 32
eval_strategy: 'no'
remove_unused_columns: true
dataset_kwargs:
  add_special_tokens: false
  append_concat_token: false
output_dir: results/sft/ali-elganzory-1.7b-Comma0.1-300BT-longsft_16k_tulu-3-sft-mixture-decontaminated_H100-masked
overwrite_output_dir: false
save_strategy: steps
save_steps: 200
save_total_limit: 2
push_to_hub: 'False'
log_level: info
report_to:
- wandb
logging_steps: 10
logging_strategy: steps
seed: 42
bf16: true
gradient_checkpointing: true
learning_rate: 5.0e-06
lr_scheduler_type: linear
max_length: 4096
warmup_ratio: 0.03
num_train_epochs: 2
per_device_train_batch_size: 16
gradient_accumulation_steps: 1
assistant_only_loss: true
hub_model_id: ali-elganzory/1.7b-Comma0.1-300BT-longsft_16k-SFT-Tulu3-decontaminated-masked
hub_strategy: every_save