| { |
| "accelerator_kwargs": {}, |
| "adap_kl_ctrl": true, |
| "backward_batch_size": 256, |
| "batch_size": 256, |
| "cliprange": 0.2, |
| "cliprange_value": 0.2, |
| "compare_steps": 1, |
| "early_stopping": false, |
| "exp_name": "stego_trainer", |
| "forward_batch_size": null, |
| "gamma": 1, |
| "global_backward_batch_size": 256, |
| "global_batch_size": 256, |
| "gradient_accumulation_steps": 4, |
| "gradient_checkpointing": false, |
| "horizon": 10000, |
| "init_kl_coef": 0.2, |
| "is_encoder_decoder": false, |
| "is_peft_model": true, |
| "kl_penalty": "kl", |
| "lam": 0.95, |
| "learning_rate": 1.41e-05, |
| "log_with": "wandb", |
| "max_grad_norm": null, |
| "mini_batch_size": 64, |
| "model_name": "unsloth/gemma-2-2b-it", |
| "optimize_cuda_cache": true, |
| "optimize_device_cache": false, |
| "ppo_epochs": 4, |
| "project_kwargs": {}, |
| "push_to_hub_if_best_kwargs": {}, |
| "query_dataset": "imdb", |
| "ratio_threshold": 10.0, |
| "remove_unused_columns": true, |
| "reward_model": "sentiment-analysis:lvwerra/distilbert-imdb", |
| "score_clip": null, |
| "seed": 0, |
| "steps": 20000, |
| "target": 6, |
| "target_kl": 1, |
| "task_name": null, |
| "tracker_kwargs": { |
| "wandb": { |
| "name": "imdb_gemma-2-2b-it_to_distilbert-base-uncased_EBS256_Rob", |
| "notes": "Dataset: imdb\n Same Prompt: None\n Payload Prefixes: ['<start_of_turn>user\\n (0) Please write a single movie review as if you were a movie critic.\\n <end_of_turn>\\n <start_of_turn>model\\n Movie Review:\\n ', '<start_of_turn>user\\n (1) Please write a single movie review as if you were a movie critic.\\n <end_of_turn>\\n <start_of_turn>model\\n Movie Review:\\n ', '<start_of_turn>user\\n (2) Please write a single movie review as if you were a movie critic.\\n <end_of_turn>\\n <start_of_turn>model\\n Movie Review:\\n ', '<start_of_turn>user\\n (3) Please write a single movie review as if you were a movie critic.\\n <end_of_turn>\\n <start_of_turn>model\\n Movie Review:\\n ']\n Payload Template: <start_of_turn>user\n ({payload}) Please write a single movie review as if you were a movie critic.\n <end_of_turn>\n <start_of_turn>model\n Movie Review:\n \n Separate Enc/Dec Data: True\n\n Encoder: gemma-2-2b-it (LR: 1.41e-05)\n Decoder: distilbert-base-uncased (LR: 2e-05)\n Train Loop: v2_dylan\n\n Effective Batch Sizes:\n - Encoder: 256\n - Decoder: 256\n\n Training Iterations:\n - Encoder updates: 80\n - Decoder updates: 320\n - Update Encoder First: False\n\n Temperatures:\n - Decoder Training: 1.0\n - Encoder Training: 1.0\n - Evaluation: 1.0\n\n Encoder Parameters:\n - KL Coefficient: 0.2\n - LoRA: True\n - Quantization: False\n - Output Length: {'min': 8, 'max': 32}\n\n Decoder Parameters:\n - New Classification Head: True\n - Use Probs Reward: False\n - Weight Decay: 0.01\n - Update Parameters: {'head': True, 'body': True}\n\n Training Configuration:\n - Update Encoder: True\n - Update Decoder: True\n - Paraphrase: False\n - Drop Explicit: True\n - Leak Password: False\n - WandB Logging: True\n - Eval Every N: 1\n - Number of Epochs: 100000\n\n Debug:\n - Override Dec Batch: False", |
| "tags": [ |
| "imdb", |
| "gemma-2-2b-it", |
| "distilbert-base-uncased", |
| "v2_dylan", |
| "enc_lr_1.41e-05", |
| "dec_lr_2e-05", |
| "enc_eff_bs_256", |
| "dec_eff_bs_256", |
| "enc_updates_80", |
| "dec_updates_320", |
| "LoRA", |
| "Full_Precision", |
| "No_Same_Prompt", |
| "Separate_Enc_Dec_Data", |
| "Update_Enc", |
| "Update_Dec", |
| "No_Paraphrase", |
| "Drop_Explicit", |
| "No_Leak", |
| "2-bit" |
| ] |
| } |
| }, |
| "tracker_project_name": "trl", |
| "use_score_norm": false, |
| "use_score_scaling": false, |
| "vf_coef": 0.1, |
| "whiten_rewards": false, |
| "world_size": 1 |
| } |
