init: add LoRA training infrastructure and 20k dataset

training/scripts/prepare_dataset.py

@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+"""
+Prepare Cyron summary dataset for LoRA training.
+
+Reads combined_20k.jsonl and formats it for Hugging Face TRL training.
+"""
+
+import json
+import argparse
+from pathlib import Path
+
+
+def prepare_dataset(input_file, output_file, test_size=0.05):
+    """Prepare dataset for training."""
+    
+    with open(input_file) as f:
+        examples = [json.loads(line) for line in f]
+    
+    print(f"Loaded {len(examples)} examples")
+    
+    # Format for training
+    formatted = []
+    for ex in examples:
+        # Create instruction-input-output format
+        instruction = f"Task: {ex['task']}"
+        if ex['files_changed']:
+            instruction += f"\nFiles: {', '.join(ex['files_changed'][:3])}"
+        
+        input_text = ""
+        if ex['tests_run']:
+            input_text += f"Tests run: {ex['test_count']}"
+        if ex['commit']:
+            input_text += f"\nCommit: {ex['git']['commit'] if ex.get('git') else 'yes'}"
+        if ex['push']:
+            input_text += "\nPush: yes"
+        
+        output_text = ex['output']
+        
+        # Create conversation format
+        conversation = {
+            "conversations": [
+                {
+                    "from": "human",
+                    "value": f"Generate a Cyron summary for this task:\n\n{instruction}\n\n{input_text}"
+                },
+                {
+                    "from": "gpt",
+                    "value": output_text
+                }
+            ]
+        }
+        
+        formatted.append(conversation)
+    
+    # Split into train/test
+    import random
+    random.seed(42)
+    random.shuffle(formatted)
+    
+    split_point = int(len(formatted) * (1 - test_size))
+    train_data = formatted[:split_point]
+    test_data = formatted[split_point:]
+    
+    # Save
+    with open(output_file.parent / "train.jsonl", "w") as f:
+        for item in train_data:
+            f.write(json.dumps(item) + "\n")
+    
+    with open(output_file.parent / "test.jsonl", "w") as f:
+        for item in test_data:
+            f.write(json.dumps(item) + "\n")
+    
+    print(f"Train: {len(train_data)} examples")
+    print(f"Test: {len(test_data)} examples")
+    print(f"Saved to {output_file.parent}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Prepare LoRA training dataset")
+    parser.add_argument("--input", type=str, default="../combined_20k.jsonl", 
+                       help="Input combined dataset")
+    parser.add_argument("--output", type=str, default="data", 
+                       help="Output directory")
+    parser.add_argument("--test-size", type=float, default=0.05, 
+                       help="Test set percentage")
+    args = parser.parse_args()
+    
+    output_dir = Path(args.output)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    
+    prepare_dataset(args.input, output_dir, args.test_size)
+
+
+if __name__ == "__main__":
+    main()