{
  "train_loop_config": {
    "_dataset": "xsum",
    "_scenario": "clamp_sft_ref",
    "a2g_norm": true,
    "base_path": "/common/home/users/m/mq.nguyen.2023/testcode/SAC_LM/module9_clmv3",
    "beta_coef": "0.1",
    "block_size": 1024,
    "buffer_max_size": 20000,
    "causal_model": true,
    "checkpoint_interval": "2_000",
    "clamp_ref": true,
    "clamp_update": false,
    "config_name": null,
    "constraint_type": "kl",
    "cpu_per_worker": 6,
    "dataset_config_name": null,
    "dataset_name": "nbtpj/summ_ds_train",
    "effective_batch_size": 16,
    "ent_coef": 0.0001,
    "epsilon": 0.2,
    "eval_interval": "2_000",
    "freeze_role2": true,
    "gamma": 0.95,
    "gpu_per_worker": 1,
    "grad_norm": 0.5,
    "gradient_accumulation_steps": 1,
    "gradient_checkpoint": false,
    "group_relative_norm": true,
    "inference_config": {
      "do_sample": true,
      "max_new_tokens": 40,
      "min_new_tokens": 15,
      "temperature": 0.0
    },
    "keep_eval_size": false,
    "label_col": "summary",
    "learning_rate": 3e-07,
    "lm_fraction": -1.0,
    "lm_w": 0.0,
    "log_interval": "5m",
    "log_rollout_txt": true,
    "lora": false,
    "lr_scheduler_type": "constant",
    "max_train_rollouts": 100000,
    "max_train_steps": 100000,
    "mini_epoch": 2,
    "mixed_precision": "bf16",
    "model_name_or_path": "gpt2",
    "model_type": null,
    "n_augment": 0,
    "n_generate": 4,
    "need_attn_mask": true,
    "no_is_correction": false,
    "num_warmup_steps": 200,
    "only_train_role1": true,
    "per_device_eval_batch_size": 16,
    "per_device_query_rollout_batch_size": 32,
    "per_device_train_batch_size": 16,
    "pretrained_role2_name_or_path": "gpt2",
    "prompt_0": "{text}",
    "prompt_1": "{text}\nTL;DR: ",
    "prompt_2": "Given the text: {role1_output}\nReconstruct the summarized text to the detailed:",
    "prompt_eval": "{text}\nTL;DR:",
    "push_to_hub": null,
    "ref_role1_name_or_path": "nbtpj/summ_gpt2_tldr_xsum",
    "ref_role2_name_or_path": "gpt2",
    "rl_algo": "on_policy",
    "rl_w": 1.0,
    "rollout_config": {
      "acc_scale": "log",
      "accuracy_w": 1000.0,
      "accuracy_w2": 0.01,
      "len_pen": 1.0,
      "len_pen2": 1.0,
      "similarity_fn": "rouge",
      "threshold": 0.002
    },
    "rollout_game": "baseline3v2",
    "sample_config": {
      "do_sample": true,
      "min_new_tokens": 5,
      "temperature": 1.0
    },
    "script": "/common/home/users/m/mq.nguyen.2023/testcode/SAC_LM/module9_clmv3/execute/role_ablation/role_ablation.py",
    "seed": 0,
    "test_clm": false,
    "test_gen": true,
    "test_glue": false,
    "text_col": "text",
    "text_template": "{text}\nTL;DR: {summary}",
    "tokenizer_name": null,
    "train_from_raw": true,
    "train_split_name": "sim_with_one_golden__xsum_train",
    "trunc_eval": 20000,
    "trunc_evals": [
      "xsum___20000"
    ],
    "trust_remote_code": true,
    "tune_metrics": [
      "xsum/rouge1___1.0",
      "xsum/rouge2___2.0",
      "xsum/bertscore_f1___0.25"
    ],
    "use_deepspeed": false,
    "use_slow_tokenizer": false,
    "vectorizer_path": "/common/home/users/m/mq.nguyen.2023/testcode/SAC_LM/module9_clmv3/vectorizer/wikitext103_tfidf_full.joblib",
    "vllm_sleep": true,
    "vllm_vram_ratio": 0.3,
    "weight_decay": 1e-05,
    "world_size": 1,
    "zero_config": 2
  }
}