Merge pull request #41 from leonvanbokhorst/model-bench

feat: Enhance prompt formatting and optimize model performance

src/04_fine_tuning.py

@@ -21,6 +21,7 @@
 
 dotenv.load_dotenv()
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
+HF_TOKEN = os.getenv("HF_TOKEN")
 
 
 def initialize_tokenizer(model_name: str, hf_token: str) -> AutoTokenizer:
@@ -156,8 +157,21 @@ def filter_quality(example: Dict[str, Any]) -> bool:
     return special_char_ratio <= 0.2
 
 
+def format_prompt(instruction: str, response: str = "") -> str:
+    """Format the prompt for the model with emphasis on positive and helpful responses."""
+    return f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+
+Cutting Knowledge Date: December 2023
+Today Date: 23 July 2024
+
+You are a helpful and polite assistant<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+Tell me about {instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+{response}<|eot_id|>"""
+
+
 def prepare_dataset(tokenizer):
-    """Prepare dataset with Alpaca-style prompt template"""
     dataset = load_dataset("leonvanbokhorst/synthetic-complaints-v2")
 
     # Use full validation set and shuffle
@@ -181,15 +195,11 @@ def format_prompt(instruction: str, response: str = "") -> str:
         return f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
 
 Cutting Knowledge Date: December 2023
-Today Date: {datetime.now().strftime('%d %b %Y')}
-
-You are a helpful AI assistant.<|eot_id|>
+Today Date: 23 July 2024
 
-<|start_header_id|>user<|end_header_id|>
+You are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>
 
-{instruction}<|eot_id|>
-
-<|start_header_id|>assistant<|end_header_id|>
+Tell me about {instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
 {response}<|eot_id|>"""
 
@@ -250,6 +260,7 @@ def inference_example(model, tokenizer, prompt: str) -> str:
     try:
         device = model.device
         formatted_prompt = format_prompt(prompt)
+        print(f"Formatted Prompt: {formatted_prompt}")
 
         model_inputs = tokenizer(
             formatted_prompt,
@@ -260,6 +271,8 @@ def inference_example(model, tokenizer, prompt: str) -> str:
             return_token_type_ids=False,
         ).to(device)
 
+        print(f"Tokenized Input IDs: {model_inputs['input_ids']}")
+
         with torch.no_grad():
             outputs = model.generate(
                 input_ids=model_inputs["input_ids"],
@@ -276,7 +289,10 @@ def inference_example(model, tokenizer, prompt: str) -> str:
                 eos_token_id=tokenizer.eos_token_id,
             )
 
+        print(f"Raw Output IDs: {outputs}")
+
         response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+        print(f"Decoded Response: {response}")
 
         # Extract only the assistant's response
         if "<|start_header_id|>assistant<|end_header_id|>" in response:
@@ -372,9 +388,9 @@ def get_training_args() -> TrainingArguments:
         # WSL2 has better memory management than Windows, allowing for more aggressive batching
         # RTX 4090 has 24GB VRAM and WSL2 can utilize it more efficiently
         # Effective batch size = per_device_batch * gradient_accumulation = 8 * 4 = 32
-        per_device_train_batch_size=8,  # Optimal for 24GB VRAM under WSL2
-        per_device_eval_batch_size=8,  # Match training batch size
-        gradient_accumulation_steps=4,  # Accumulate for larger effective batch
+        per_device_train_batch_size=12,  # Optimal for 24GB VRAM under WSL2
+        per_device_eval_batch_size=12,  # Match training batch size
+        gradient_accumulation_steps=6,  # Accumulate for larger effective batch
         # Data Loading Optimization
         # WSL2's Linux kernel provides better process management than Windows
         # Can use more CPU cores efficiently without system instability

src/11_llm_benchmark.py

@@ -65,15 +65,13 @@ def __init__(
         self.model = AutoModelForCausalLM.from_pretrained(
             model_name,
             device_map="auto",
-            token=HF_TOKEN,
             torch_dtype=torch.float16,
         ).to(
             self.device
         )  # Explicitly move to device
 
         self.tokenizer = AutoTokenizer.from_pretrained(
             model_name,
-            token=HF_TOKEN,
         )
 
         # Load reference model for comparison
@@ -120,8 +118,22 @@ def generate_response(
         - top_p=0.9: Nucleus sampling for natural language variation
         - do_sample=True: Enables probabilistic sampling
         """
+        # Use the new prompt template
+        formatted_prompt = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+
+Cutting Knowledge Date: December 2023
+Today Date: {datetime.now().strftime('%d %b %Y')}
+
+You are a helpful assistant <|eot_id|>
+
+<|start_header_id|>user <|end_header_id|>
+
+{prompt} <|eot_id|>
+<|start_header_id|>assistant <|end_header_id|>
+"""
+
         inputs = tokenizer(
-            prompt,
+            formatted_prompt,
             return_tensors="pt",
             padding=True,
             truncation=True,
@@ -273,10 +285,20 @@ def calculate_metrics(
 
         return metrics
 
+    def clear_model_caches(self):
+        """
+        Clear model caches to free up memory.
+        """
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
     def run_benchmark(self, num_samples: int = 100) -> Dict[str, Any]:
         """
         Run comprehensive benchmark comparing models with detailed progress reporting.
         """
+        # Clear caches before starting the benchmark
+        self.clear_model_caches()
+
         dataset = load_dataset("leonvanbokhorst/synthetic-complaints-v2")["train"]
         test_samples = dataset.select(range(num_samples))
 
@@ -379,8 +401,8 @@ def run_benchmark(self, num_samples: int = 100) -> Dict[str, Any]:
                 # Print sample outputs
                 tqdm.write("\nSample Outputs:")
                 tqdm.write("Topic: " + sample["topic"])
-                tqdm.write("Fine-tuned (first 100 chars): " + ft_response[:100] + "...")
-                tqdm.write("Reference (first 100 chars): " + ref_response[:100] + "...")
+                tqdm.write("Fine-tuned (first 256 chars): " + ft_response[:256])
+                tqdm.write("Reference (first 256 chars): " + ref_response[:256])
                 tqdm.write("\n" + "=" * 50)
 
             # Print warning if metrics are concerning

src/poc/fine_tune_test.py

@@ -21,7 +21,7 @@
     tokenizer.pad_token_id = tokenizer.eos_token_id
 
 # Load the model with exact same settings as training
-device = "mps" if torch.backends.mps.is_available() else "cpu"
+device = "mps" if torch.backends.mps.is_available() else "cuda" if torch.cuda.is_available() else "cpu"
 print(f"🖥️ Using device: {device}")
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
@@ -36,15 +36,19 @@
 print("Model type:", type(model))
 
 # Add this before the generate_complaint function
-PROMPT_TEMPLATE = """### Instruction:
-{0}
+PROMPT_TEMPLATE = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
 
-### Response:
+Cutting Knowledge Date: December 2023
+Today Date: 23 July 2024
+
+You are a helpful assistant<|eot_id|><|start_header_id|>user <|end_header_id|>
+
+{0}<|eot_id|><|start_header_id|>assistant <|end_header_id|>
 """
 
 def generate_complaint(prompt: str) -> str:
     # Make the instruction more explicit
-    instruction = f"Write an angry complaint about {prompt}. Be specific about why you are upset and express your frustration clearly."
+    instruction = f"Tell me about {prompt}."
 
     formatted_prompt = PROMPT_TEMPLATE.format(instruction)