| openai_base_url: "https://openrouter.ai/api/v1/" | |
| student_model_name: "meta-llama/Llama-3.2-3B-Instruct" | |
| teacher_model_name: "deepseek/deepseek-r1-distill-qwen-14b" | |
| dataset_name: "open-r1/OpenR1-Math-220k" | |
| output_dir: "results/llama/" | |
| save_steps: 50 | |
| learning_rate: 5.0e-07 | |
| max_new_tokens: 3072 | |
| max_feedback_new_tokens: 4096 | |
| num_return_sequences: 4 | |
| seed: 12345 | |
| temperature: 0.7 | |
| top_p: 0.9 | |
| top_k: 50 | |
| max_seq_length: 4096 | |
| cache_dir: "~/.cache" | |
| warmup_steps: 100 | |
| total_steps: 1000 | |
| max_grad_norm: 0.1 | |
| grpo_beta: 0.05 | |
| sft_beta: 0.05 | |
| thought_process_weight: 0.07 | |
| answer_weight: 0.1 | |
| format_weight: 0.03 |