Upload 3 files

train/generate_data.py

@@ -0,0 +1,38 @@
+import json
+import random
+
+def generate_disaster_data(num_samples=100):
+    dataset = []
+    grid_size = 5
+    
+    for _ in range(num_samples):
+        # Randomize positions
+        start = (random.randint(0, 4), random.randint(0, 4))
+        target = (random.randint(0, 4), random.randint(0, 4))
+        while target == start:
+            target = (random.randint(0, 4), random.randint(0, 4))
+            
+        obstacle = (random.randint(0, 4), random.randint(0, 4))
+        while obstacle == start or obstacle == target:
+            obstacle = (random.randint(0, 4), random.randint(0, 4))
+
+        instruction = (
+            f"You are a disaster response drone. Grid is {grid_size}x{grid_size}. "
+            f"Start: {start}. Goal: {target}. Obstacle at {obstacle}. "
+            "Output the move sequence (North, South, East, West) to reach the goal safely."
+        )
+        
+        # Simple placeholder logic for 'ideal' output
+        # In a real run, GRPO will learn to improve this
+        dataset.append({
+            "instruction": instruction,
+            "input": "",
+            "output": "Thinking: I must calculate the path... Final Move: [Sequence]"
+        })
+
+    with open("synthetic_data.json", "w") as f:
+        json.dump(dataset, f, indent=4)
+    print(f"✅ Successfully generated {num_samples} samples in train/synthetic_data.json")
+
+if _name_ == "_main_":
+    generate_disaster_data(200) # Generates 200 scenarios

train/grpo_trainer.py

@@ -0,0 +1,53 @@
+from unsloth import FastLanguageModel, PatchFastRL
+PatchFastRL("GRPO", FastLanguageModel)
+from trl import GRPOTrainer, GRPOConfig
+from datasets import load_dataset
+import torch
+
+# 1. Load Model & Tokenizer
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = "unsloth/meta-llama-3.1-8b-instruct-bnb-4bit",
+    max_seq_length = 512,
+    load_in_4bit = True,
+    fast_inference = True,
+)
+
+# 2. Add LoRA Adapters for training
+model = FastLanguageModel.get_peft_model(
+    model,
+    r = 16,
+    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj"],
+    lora_alpha = 16,
+    lora_dropout = 0,
+)
+
+# 3. Load your freshly generated data
+dataset = load_dataset("json", data_files="synthetic_data.json", split="train")
+
+# 4. Define Reward Functions (The logic for your Disaster Grid)
+def reward_reach_goal(compts, **kwargs):
+    # Reward for including the correct target coordinates in the output
+    rewards = [1.0 if "Final Move:" in c else 0.0 for c in compts]
+    return rewards
+
+# 5. Training Configuration
+training_args = GRPOConfig(
+    learning_rate = 5e-6,
+    num_train_epochs = 1,
+    per_device_train_batch_size = 1,
+    gradient_accumulation_steps = 4,
+    outputs_dir = "outputs",
+    optim = "adamw_8bit",
+)
+
+# 6. Initialize Trainer
+trainer = GRPOTrainer(
+    model = model,
+    reward_funcs = [reward_reach_goal],
+    args = training_args,
+    train_dataset = dataset,
+)
+
+# 7. Start Training
+trainer.train()
+model.save_pretrained_merged("disaster_model_final", tokenizer, save_method = "merged_16bit")