# ORPO iter-2 on planner-COLLAB-iter1 — starts from prior ORPO checkpoint, uses new rollouts.
model_name_or_path: ./output/qwen-coder3b-scaleup-planner-COLLAB-orpo
torch_dtype: bfloat16
use_flash_attention_2: false

dataset_mixer:
  ../data/llm_alignment/scaleup_iter2/hf_planner_collaborative: 1.0

dataset_splits:
- train_dpo
- test_dpo
preprocessing_num_workers: 12
chat_template: "{{'<|im_start|>user\n' + messages['prompt'] + '<|im_end|>\n'}}{{'<|im_start|>assistant\n' + messages['completion'] + '<|im_end|>\n'}}"
report_to: ["tensorboard"]

bf16: true
beta: 1.0
do_eval: true
eval_strategy: "steps"
eval_steps: 100
gradient_accumulation_steps: 16
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: False
learning_rate: 1.0e-6
log_level: info
logging_steps: 10
lr_scheduler_type: inverse_sqrt
max_length: 4000
max_prompt_length: 3500
num_train_epochs: -1
max_steps: 200
optim: adamw_torch
output_dir: output/qwen-coder3b-scaleup-planner-COLLAB-orpo-iter2
overwrite_output_dir: true
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
push_to_hub: false
remove_unused_columns: false
save_strategy: "steps"
save_steps: 200
save_total_limit: 1
seed: 42
warmup_ratio: 0.1
warmup_steps: 50
max_grad_norm: 0.5