| mode: train | |
| experiment: | |
| dataset_size: 6000 | |
| 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}_${algorithm.name}_${algorithm.training.curriculum_schedule}_${algorithm.training.scheduler_params.mu_exp}_${algorithm.training.scheduler_params.sigma}_${algorithm.training.scheduler_params.min_prob}_${algorithm.training.max_steps} | |
| 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: meta-llama | |
| trim: Llama-3.2-3B-Instruct | |
| name: ${model.family}/${model.trim} | |
| trust_remote_code: true | |
| torch_dtype: bfloat16 | |
| attn_implementation: flash_attention_2 | |
| task: | |
| name: blocksworld1246 | |
| data_files: | |
| - data/blocksworld/train_set-1-complete-correct.json | |
| - data/blocksworld/train_set-2-complete-correct.json | |
| - data/blocksworld/train_set-4-complete-correct.json | |
| - data/blocksworld/train_set-6-complete-correct.json | |
| icl_examples_file: data/blocksworld/train_set-2-more_with_trace.json | |
| use_icl_examples: false | |
| training: | |
| max_prompt_length: 1600 | |
| max_completion_length: 512 | |
| inference: | |
| checkpoint: 1200 | |
| steps: 4 | |
| temperature: 0.0 | |
| sc_num: 1 | |
| use_icl: false | |
| icl_num: 2 | |
| prompt_path: prompts/blocksworld/pool_prompt_v1.json | |
| data_path: data/blocksworld/split_v1/split_v1_step_{steps}_data.json | |
| config_file: data/blocksworld/bw_config.yaml | |
| domain_file: data/blocksworld/generated_domain.pddl | |
| pass_at_k: 1 | |
| num_shot: 4 | |
| resume: 0 | |
| max_new_tokens: 512 | |
| max_batch_size: 64 | |
| algorithm: | |
| name: sgrpo | |
| training: | |
| learning_rate: 1.0e-06 | |
| lr_scheduler_type: cosine | |
| logging_steps: 10 | |
| max_steps: 300 | |
| curriculum: true | |
| curriculum_schedule: classic | |
| scheduler_params: | |
| mu_exp: 0.5 | |
| sigma: 0.5 | |
| min_prob: true | |
| 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: 0.5 | |
| report_to: | |
| - wandb | |
| push_to_hub: true | |
| save_strategy: steps | |
| save_steps: ${algorithm.training.max_steps} | |
| eval_strategy: steps | |
| tf32: true | |