| mode: train | |
| experiment: | |
| dataset_size: 0 | |
| dataset_seed: 1234 | |
| test_size: 0.1 | |
| hf_token: ${oc.env:HF_TOKEN,null} | |
| output: | |
| root_path: ${oc.env:ROOT_PATH} | |
| run_name: ${model.trim}_${task.name}_${algorithm.name} | |
| lora: | |
| r: 32 | |
| alpha: 64 | |
| dropout: 0.1 | |
| target_modules: | |
| - q_proj | |
| - v_proj | |
| task_type: CAUSAL_LM | |
| occupy_gpu_memory: false | |
| occupy_gpu_memory_gb: 50 | |
| gpu_device: cuda:0 | |
| model: | |
| family: Qwen | |
| trim: Qwen2.5-0.5B-Instruct | |
| name: ${model.family}/${model.trim} | |
| trust_remote_code: true | |
| torch_dtype: bfloat16 | |
| attn_implementation: flash_attention_2 | |
| task: | |
| name: countdown34 | |
| data_files: | |
| - citrinegui/countdown_n3t100_1-100 | |
| - citrinegui/countdown_n4t100_1-100 | |
| test_file: citrinegui/countdown_n4t100_1-100 | |
| force_redownload: false | |
| train_size: 327680 | |
| test_size: 1024 | |
| num_operands: 6 | |
| max_target: 1000 | |
| min_number: 1 | |
| max_number: 100 | |
| template_type: qwen-instruct | |
| training: | |
| max_prompt_length: 1000 | |
| max_completion_length: 256 | |
| inference: | |
| checkpoint: 300 | |
| temperature: 0.7 | |
| sc_num: 1 | |
| resume: 0 | |
| max_new_tokens: 256 | |
| algorithm: | |
| name: grpo | |
| training: | |
| learning_rate: 1.0e-06 | |
| lr_scheduler_type: cosine | |
| logging_steps: 10 | |
| max_steps: 300 | |
| per_device_train_batch_size: 2 | |
| gradient_accumulation_steps: 4 | |
| gradient_checkpointing: true | |
| bf16: true | |
| num_generations: 8 | |
| beta: 0.001 | |
| use_vllm: true | |
| vllm_gpu_memory_utilization: 0.2 | |
| report_to: | |
| - wandb | |
| push_to_hub: true | |
| save_strategy: steps | |
| save_steps: 50 | |
| eval_strategy: steps | |