add-missing-config

examples/xpuyu_usage/bootcamp_rl/configs/example_training_config.py

@@ -0,0 +1,70 @@
+# Automatic Generated Configs
+
+# Model Related Settings
+actor="actor_model_path"
+reference="actor_model_path"
+chat_template="r1"
+dtype="auto"
+selective_recompute=1.0
+cpu_offload=False
+cuda_graph=True
+tp_size=4
+sp_size=1
+
+# Dataset Related Settings
+datasets="your dataset path"
+data_difficulty_balance_cfg=None
+num_workers=0
+
+# Generate Related Settings
+gen_global_batch=128
+gen_max_new=8192
+gen_max_length=10240
+gen_top_k=0
+gen_top_p=0.9
+temperature=1
+gen_do_sample=True
+max_prefill_batch=8
+prompt_repeat_k=4
+
+# Optimizer Related Settings
+rl_global_batch=128
+rl_micro_batch=4
+rl_mini_batch_steps=1
+warmup_steps=0
+actor_lr=1e-6
+actor_min_lr=1e-6
+wd=0.01
+max_grad_norm=1
+
+# General Settings
+work_dir="examples/xpuyu_usage/ckpts/experiment_name"
+checkpoint_interval=40
+log_interval=1
+seed=0
+debug=True
+
+# Reward Settings
+reward_shaping_type="grpo"
+loss_type="per_token"
+kl_coef=0.01
+stop_word="<｜end▁of▁sentence｜>"
+
+
+judgers_config = dict(
+    bootcamp_judger=dict(  
+        stop_word=stop_word,
+        num_processes=8,
+        concurrency_per_proc=(8, 8),
+    ),
+)
+
+import json
+data_judger_mapping = {}
+
+with open("dataset_path", "r") as f:
+    for line in f:
+        item = json.loads(line)
+        data_judger_mapping[item["metadata"]["data_source"]] = ["bootcamp_judger"]
+
+