55 lines
1.9 KiB
YAML
55 lines
1.9 KiB
YAML
|
model_name_or_path: ali-elganzory/1.7b-Comma0.1-300BT-longsft_16k-SFT-Tulu3-decontaminated-masked
|
||
|
|
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\" + message[\"content\"]\
|
||
|
|
\ + eos_token + \"\\n\" -}}\n\t\t{%- else -%}\n\t\t\t{{- \"<|assistant|>\\n\" +\
|
||
|
|
\ message[\"content\"] + eos_token -}}\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 -%}"
|
||
|
|
additional_special_tokens:
|
||
|
|
- <|system|>
|
||
|
|
- <|user|>
|
||
|
|
- <|assistant|>
|
||
|
|
dataset_mixture:
|
||
|
|
datasets:
|
||
|
|
- id: ali-elganzory/llama-3.1-tulu-3-8b-preference-mixture-decontaminated
|
||
|
|
config: default
|
||
|
|
split: train
|
||
|
|
columns:
|
||
|
|
- chosen
|
||
|
|
- rejected
|
||
|
|
dataset_num_proc: 32
|
||
|
|
remove_unused_columns: true
|
||
|
|
eval_strategy: 'no'
|
||
|
|
output_dir: results/dpo/ali-elganzory-1.7b-Comma0.1-300BT-longsft_16k-SFT-Tulu3-decontaminated-masked_tulu-3-8b-preference-mixture-decontaminated_H100
|
||
|
|
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
|
||
|
|
gradient_checkpointing_kwargs:
|
||
|
|
use_reentrant: false
|
||
|
|
learning_rate: 5.0e-07
|
||
|
|
lr_scheduler_type: linear
|
||
|
|
max_length: 2048
|
||
|
|
beta: 5
|
||
|
|
warmup_ratio: 0.1
|
||
|
|
per_device_train_batch_size: 8
|
||
|
|
gradient_accumulation_steps: 1
|
||
|
|
num_train_epochs: 1
|
||
|
|
hub_model_id: ali-elganzory/1.7b-Comma0.1-300BT-longsft_16k-DPO-Tulu3-decontaminated-masked
|
||
|
|
hub_strategy: every_save
|