mirror of
https://github.com/JamesTheGiblet/BuddAI.git
synced 2026-01-08 21:58:40 +00:00
f9fd27d228
- Added `ModelFineTuner` class for preparing training data and fine-tuning models based on user corrections. - Introduced `CodeValidator` class to validate generated code against various hardware and style rules, including safety checks and function naming conventions. - Developed skills for calculator operations, system information retrieval, weather fetching, and timer functionality. - Implemented a self-diagnostic skill to run unit tests and report results. - Created a dynamic skill loading mechanism to discover and register skills from the current directory. - Added unit tests for skills to ensure functionality and reliability.
41 lines
No EOL
1.3 KiB
Python
41 lines
No EOL
1.3 KiB
Python
import sqlite3
|
|
import json
|
|
from core.buddai_shared import DB_PATH, DATA_DIR
|
|
|
|
class ModelFineTuner:
|
|
"""Fine-tune local model on YOUR corrections"""
|
|
|
|
def prepare_training_data(self):
|
|
"""Convert corrections to training format"""
|
|
conn = sqlite3.connect(DB_PATH)
|
|
cursor = conn.cursor()
|
|
|
|
cursor.execute("""
|
|
SELECT original_code, corrected_code, reason
|
|
FROM corrections
|
|
""")
|
|
|
|
training_data = []
|
|
for original, corrected, reason in cursor.fetchall():
|
|
training_data.append({
|
|
"prompt": f"Generate code for: {reason}",
|
|
"completion": corrected,
|
|
"negative_example": original
|
|
})
|
|
|
|
conn.close()
|
|
|
|
# Save as JSONL for fine-tuning
|
|
output_path = DATA_DIR / 'training_data.jsonl'
|
|
with open(output_path, 'w', encoding='utf-8') as f:
|
|
for item in training_data:
|
|
f.write(json.dumps(item) + '\n')
|
|
return f"Exported {len(training_data)} examples to {output_path}"
|
|
|
|
def fine_tune_model(self):
|
|
"""Fine-tune Qwen on your corrections"""
|
|
# This requires:
|
|
# 1. Export training data
|
|
# 2. Use Ollama modelfile or external training
|
|
# 3. Create custom model: qwen2.5-coder-james:3b
|
|
pass |