| { | |
| "d_model": 512, | |
| "n_heads": 8, | |
| "n_kv_heads": 2, | |
| "head_dim": 64, | |
| "n_layers": 4, | |
| "loop_count": 2, | |
| "max_loops": 16, | |
| "max_seq_len": 512, | |
| "rope_theta": 10000.0, | |
| "loop_theta": 100.0, | |
| "mlp_ratio": 2.6666666666666665, | |
| "sft_data_path": "/content/computer_sft.jsonl", | |
| "sft_checkpoint": "computer_sft_final.pt", | |
| "sft_batch_size": 8, | |
| "sft_grad_accum": 4, | |
| "sft_epochs": 8, | |
| "sft_lr": 3e-05, | |
| "weight_decay": 0.05, | |
| "grad_clip": 1.0, | |
| "rl_checkpoint": "computer_rl_final.pt", | |
| "rl_epochs": 30, | |
| "rl_lr": 2e-06, | |
| "entropy_coef": 0.002, | |
| "baseline_beta": 0.9, | |
| "max_episode_steps": 8, | |
| "policy_temperature": 0.7, | |
| "step_penalty": -0.01, | |
| "progress_reward": 0.3, | |
| "success_reward": 2.0, | |
| "wrong_action_penalty": -0.2, | |
| "repeat_penalty": -0.15, | |
| "premature_done_penalty": -0.5, | |
| "terminal_wrong_penalty": -0.3, | |
| "pretrained_checkpoint": "checkpoint.pt", | |
| "seed": 42, | |
| "architectures": [ | |
| "LoopTransformerLM" | |
| ], | |
| "model_type": "loop_gqa_action", | |
| "vocab_size": 72000, | |
| "tokenizer_name": "j-llm/j72-tokenizer", | |
| "auto_map": { | |
| "AutoConfig": "configuration_loop_gqa_action.LoopGQAActionConfig", | |
| "AutoModelForCausalLM": "modeling_loop_gqa_action.LoopGQAActionModel" | |
| } | |
| } |