fix: reduce grpo training runtime

Use a smaller hackathon training run with fewer episodes, generations, and completion tokens so the A100 Space can finish in a practical window.

Made-with: Cursor

training/train_grpo.py

@@ -29,18 +29,18 @@ LOAD_IN_4BIT = True
 LORA_R = 16
 LORA_TARGET_MODULES = ["q_proj", "k_proj", "v_proj", "o_proj"]
 
-NUM_TRAIN_EPOCHS = 3
+NUM_TRAIN_EPOCHS = 1
 BATCH_SIZE = 4
 GRAD_ACCUM = 4
 LR = 5e-5
-MAX_COMPLETION_LENGTH = 512
-NUM_GENERATIONS = 8
+MAX_COMPLETION_LENGTH = 128
+NUM_GENERATIONS = 4
 LOGGING_STEPS = 10
-SAVE_STEPS = 50
+SAVE_STEPS = 25
 SEED = 42
 
-N_TRAIN_EPISODES = 500
-N_EVAL_EPISODES = 50
+N_TRAIN_EPISODES = 100
+N_EVAL_EPISODES = 20
 
 WANDB_PROJECT = "context-corruption-env"
 WANDB_RUN_NAME = "qwen-1.5b-grpo-run1"