59 lines
1.1 KiB
YAML
59 lines
1.1 KiB
YAML
base_model: Qwen/Qwen2.5-3B
|
|
# optionally might have model_type or tokenizer_type
|
|
model_type: AutoModelForTokenClassification
|
|
num_labels: 2
|
|
tokenizer_type: AutoTokenizer
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
process_reward_model: true
|
|
chat_template:
|
|
datasets:
|
|
- path: trl-lib/math_shepherd
|
|
type: stepwise_supervised
|
|
step_separator: "\n"
|
|
max_completion_length:
|
|
train_on_last_step_only: false
|
|
|
|
val_set_size: 0.2
|
|
output_dir: ./outputs/out
|
|
remove_unused_columns: false
|
|
|
|
sequence_len: 2048
|
|
sample_packing: false
|
|
eval_sample_packing: false
|
|
pad_to_sequence_len: true
|
|
|
|
wandb_project:
|
|
wandb_entity:
|
|
wandb_watch:
|
|
wandb_name:
|
|
wandb_log_model:
|
|
|
|
|
|
gradient_accumulation_steps: 1
|
|
micro_batch_size: 8
|
|
eval_batch_size: 8
|
|
num_epochs: 1
|
|
optimizer: adamw_torch
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.0002
|
|
|
|
bf16: true
|
|
fp16:
|
|
tf32:
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: false
|
|
resume_from_checkpoint:
|
|
logging_steps: 1
|
|
attention: flash
|
|
|
|
|
|
warmup_ratio: 0.1
|
|
evals_per_epoch:
|
|
eval_steps: 100
|
|
saves_per_epoch: 1
|
|
weight_decay: 0.0
|
|
special_tokens:
|