axolotl version: 0.4.1
base_model: winglian/m12b-20240721-test010
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: false
strict: false
chat_template: chatml
rl: simpo
rl_beta: 2.5
cpo_alpha: 0.05
simpo_gamma: 0.1
datasets:
- path: princeton-nlp/gemma2-ultrafeedback-armorm
type: chat_template.default
chat_template: chatml
field_messages: chosen
field_chosen: chosen
field_rejected: rejected
message_field_role: role
message_field_content: content
roles:
system:
- system
user:
- user
assistant:
- assistant
dataset_prepared_path:
val_set_size: 0.0
output_dir: ./outputs/simpo-out
save_safetensors: true
save_only_model: true # fsdp seems to crap out saving the optimizer
sequence_len: 8192
sample_packing: false
pad_to_sequence_len: true
adapter:
lora_model_dir:
lora_r: 256
lora_alpha: 256
lora_dropout: 0.1
lora_target_linear: true
lora_fan_in_fan_out:
# peft_use_rslora: true
wandb_project: romulus-12b
wandb_entity: oaaic
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 5.0e-7
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:
warmup_steps: 25
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
weight_decay: 0.0
fsdp:
fsdp_config:
outputs/simpo-out
This model is a fine-tuned version of winglian/m12b-20240721-test010 on an unknown dataset.
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: 5e-07
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 16
- total_train_batch_size: 128
- total_eval_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 25
- training_steps: 466
Training results
Framework versions
- Transformers 4.43.1
- Pytorch 2.3.1+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1