初始化项目，由ModelHub XC社区提供模型

Model: divelab/DAPO_E2H-math-gaussian_0p5_0p5
Source: Original Platform

.hydra/config.yaml

@@ -0,0 +1,88 @@
+mode: train
+experiment:
+  dataset_size: 6000
+  dataset_seed: 1234
+  test_size: 0.1
+  hf_token: ${oc.env:HF_TOKEN,null}
+output:
+  root_path: ${oc.env:ROOT_PATH}
+  run_name: ${model.trim}_${task.name}_${algorithm.name}_${algorithm.training.curriculum_schedule}_${algorithm.training.scheduler_params.mu_exp}_${algorithm.training.scheduler_params.sigma}_SEC${algorithm.training.scheduler_params.vrex_adds.sec}DRO${algorithm.training.scheduler_params.vrex_adds.groupdro}G${algorithm.training.scheduler_params.vrex_adds.gaussian}_minp${algorithm.training.scheduler_params.min_prob}${ckpt2short:${algorithm.training.resume_from_checkpoint}}_${algorithm.training.max_steps}
+lora:
+  r: 32
+  alpha: 64
+  dropout: 0.1
+  target_modules:
+  - q_proj
+  - v_proj
+  task_type: CAUSAL_LM
+occupy_gpu_memory: false
+occupy_gpu_memory_gb: 50
+gpu_device: cuda:0
+model:
+  family: Qwen
+  trim: Qwen2.5-1.5B-Instruct
+  name: ${model.family}/${model.trim}
+  trust_remote_code: true
+  torch_dtype: bfloat16
+  attn_implementation: flash_attention_2
+task:
+  name: math
+  data_files:
+  - data/math/level_1
+  - data/math/level_2
+  - data/math/level_3
+  - data/math/level_4
+  training:
+    max_prompt_length: 1600
+    max_completion_length: 1600
+  inference:
+    data_files:
+    - data/math/level1
+    - data/math/level2
+    - data/math/level3
+    - data/math/level4
+    - data/math/level5
+    max_prompt_length: 1600
+    max_completion_length: 1600
+    temperature: 0.0
+    'n': 1
+algorithm:
+  name: grpo
+  training:
+    resume_from_checkpoint: null
+    learning_rate: 1.0e-06
+    lr_scheduler_type: cosine
+    logging_steps: 10
+    max_steps: 1600
+    per_device_train_batch_size: 16
+    generation_batch_size: null
+    steps_per_generation: 1
+    gradient_accumulation_steps: 4
+    gradient_checkpointing: true
+    bf16: true
+    report_to:
+    - wandb
+    push_to_hub: true
+    save_strategy: steps
+    save_steps: ${algorithm.training.max_steps}
+    tf32: true
+    num_generations: 8
+    beta: 0.001
+    use_vllm: true
+    vllm_mode: colocate
+    vllm_gpu_memory_utilization: 0.25
+    vllm_server_port: 8000
+    curriculum: false
+    curriculum_schedule: gaussian
+    scheduler_params:
+      mu_exp: 0.5
+      sigma: 0.5
+      vrex_adds:
+        groupdro: 1.0
+        gaussian: 0.0
+        sec: 0.3
+      beta: 1.0
+      min_prob: true
+      td_alpha: 0.5
+      sec_temperature: 0.3
+      max_dapo_iter: 2