Llama-3.2-3B-Code-Knowledge-Value-Eval-lora

This model is a fine-tuned version of meta-llama/Llama-3.2-3B on the kimsan0622/code-knowledge-eval dataset. It achieves the following results on the evaluation set:

• Loss: 0.9173
• Accuracy: 0.5945

Model Description

The model trained on the Code Knowledge Value Evaluation Dataset is designed to assess the educational and knowledge value of code snippets. It leverages patterns and contextual information from a large collection of open-source code, sourced from the bigcode/the-stack repository. By analyzing these code samples, the model can evaluate their utility in teaching coding concepts, solving problems, and improving developer education.

The model focuses on understanding the structure, syntax, and logic of various programming languages, enabling it to provide insights into the learning potential and technical depth of different code samples. The dataset used for training consists of 22,786 samples for training, 4,555 for validation, and 18,232 for testing, ensuring that the model is both robust and well-generalized across different coding contexts.

Intended Uses & Limitations

Intended Uses:

1. Automated Code Review: The model can be applied in automated systems to assess the knowledge value of code during code review processes. It can help identify areas where code could be optimized for better readability, maintainability, and educational impact.
2. Educational Feedback: For instructors and educational platforms, the model can offer feedback on the effectiveness of code samples used in teaching, helping to improve curriculum materials and select code that best conveys core programming concepts.
3. Curriculum Development: The model can aid in designing coding courses or instructional materials by suggesting code examples that have higher educational value, supporting a more effective learning experience.
4. Technical Skill Assessment: Organizations or platforms can use the model to assess the complexity and educational value of code submissions in coding challenges or exams.

Limitations:

1. Narrow Scope in Knowledge Evaluation: The model is specialized in evaluating code from an educational standpoint, focusing primarily on learning potential rather than production-level code quality (e.g., performance optimization or security).
2. Language and Domain Limitations: Since the dataset is sourced from bigcode/the-stack, it may not cover all programming languages or specialized domains. The model may perform less effectively in underrepresented languages or niche coding styles not well-represented in the dataset.
3. Not Suitable for All Educational Levels: While the model is designed to evaluate code for educational purposes, its outputs may be better suited for certain levels (e.g., beginner or intermediate coding), and its recommendations might not fully cater to advanced or highly specialized learners.

How to use this model?

import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel, PeftConfig

# Define the model name or path for loading the tokenizer and model using LoRA fine-tuning
model_name_or_path = "kimsan0622/Llama-3.2-3B-Code-Knowledge-Value-Eval-lora"

# Load the PEFT (Parameter-Efficient Fine-Tuning) configuration from the pretrained model
config = PeftConfig.from_pretrained(model_name_or_path)

# Load the base model for sequence classification, setting up for 6 possible labels
inference_model = AutoModelForSequenceClassification.from_pretrained(
    config.base_model_name_or_path,  # Base model path
    device_map="cuda:0",  # Use the first CUDA device for inference
    label2id={str(k): k for k in range(6)},  # Map label names (0-5) to IDs
    id2label={k: str(k) for k in range(6)},  # Map label IDs to names (0-5)
    num_labels=6,  # Define the number of labels for classification (0 to 5)
)

# Load the tokenizer for the base model
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Set the padding token if it is not already defined, matching it with the EOS token
if not tokenizer.pad_token_id: 
    tokenizer.pad_token_id = tokenizer.eos_token_id
    inference_model.config.pad_token_id = inference_model.config.eos_token_id

# Load the PEFT model using the pre-trained LoRA model and the base model
model = PeftModel.from_pretrained(inference_model, model_name_or_path)

# Sample code input to evaluate
code = [
"""
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel, PeftConfig

# Define the model name or path for loading the tokenizer and model using LoRA fine-tuning
model_name_or_path = "kimsan0622/Llama-3.2-3B-Code-Knowledge-Value-Eval-lora"

# Load the PEFT (Parameter-Efficient Fine-Tuning) configuration from the pretrained model
config = PeftConfig.from_pretrained(model_name_or_path)

# Load the base model for sequence classification, setting up for 6 possible labels
inference_model = AutoModelForSequenceClassification.from_pretrained(
    config.base_model_name_or_path,  # Base model path
    device_map="cuda:0",  # Use the first CUDA device for inference
    label2id={str(k): k for k in range(6)},  # Map label names (0-5) to IDs
    id2label={k: str(k) for k in range(6)},  # Map label IDs to names (0-5)
    num_labels=6,  # Define the number of labels for classification (0 to 5)
)

# Load the tokenizer for the base model
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Set the padding token if it is not already defined, matching it with the EOS token
if not tokenizer.pad_token_id: 
    tokenizer.pad_token_id = tokenizer.eos_token_id
    inference_model.config.pad_token_id = inference_model.config.eos_token_id

# Load the PEFT model using the pre-trained LoRA model and the base model
model = PeftModel.from_pretrained(inference_model, model_name_or_path)

# Sample code input to evaluate
code = ["code"]

# Tokenize the input code, setting the maximum length and ensuring proper padding and truncation
batch = tokenizer(code, max_length=1024, padding=True, truncation=True, return_tensors="pt")

# Perform inference without computing gradients for faster processing
with torch.no_grad():
    # Pass the input IDs and attention mask to the model for prediction
    res = model(
            input_ids=batch["input_ids"].to("cuda:0"),
            attention_mask=batch["attention_mask"].to("cuda:0"),
        )
    
    # Move the logits to the CPU and convert them to a numpy array
    preds = res.logits.cpu().numpy()
    
    # Get the predicted label by taking the argmax of the logits
    preds = np.argmax(preds, axis=1).tolist()
    
    # Print the predicted labels
    print(preds)
"""
]

# Tokenize the input code, setting the maximum length and ensuring proper padding and truncation
batch = tokenizer(code, max_length=1024, padding=True, truncation=True, return_tensors="pt")

# Perform inference without computing gradients for faster processing
with torch.no_grad():
    # Pass the input IDs and attention mask to the model for prediction
    res = model(
            input_ids=batch["input_ids"].to("cuda:0"),
            attention_mask=batch["attention_mask"].to("cuda:0"),
        )
    
    # Move the logits to the CPU and convert them to a numpy array
    preds = res.logits.cpu().numpy()
    
    # Get the predicted label by taking the argmax of the logits
    preds = np.argmax(preds, axis=1).tolist()
    
    # Print the predicted labels
    print(preds)

8 Bit quantization

from transformers import AutoTokenizer, AutoModelForSequenceClassification, BitsAndBytesConfig
from peft import PeftModel, PeftConfig

# Define the model name or path for loading the LoRA fine-tuned model
model_name_or_path = "kimsan0622/Llama-3.2-3B-Code-Knowledge-Value-Eval-lora"

# Configure the model to load in 8-bit precision to optimize memory usage and speed
bnb_config = BitsAndBytesConfig(load_in_8bit=True)

# Load the PEFT (Parameter-Efficient Fine-Tuning) configuration from the pre-trained model
config = PeftConfig.from_pretrained(model_path)

# Load the base model for sequence classification with 8-bit quantization and a device map to the first CUDA device
inference_model = AutoModelForSequenceClassification.from_pretrained(
    config.base_model_name_or_path,   # Base model path from PEFT config
    quantization_config=bnb_config,   # Apply 8-bit quantization for memory efficiency
    device_map="cuda:0",              # Map the model to the first CUDA device
    label2id={str(k): k for k in range(6)},  # Map label names (0-5) to label IDs
    id2label={k: str(k) for k in range(6)},  # Map label IDs to label names (0-5)
    num_labels=6,  # Specify the number of labels for classification
)

# Load the tokenizer associated with the base model
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Set the padding token if it's not defined, using the EOS token as the fallback
if not tokenizer.pad_token_id: 
    tokenizer.pad_token_id = tokenizer.eos_token_id
    inference_model.config.pad_token_id = inference_model.config.eos_token_id

# Load the PEFT model by applying LoRA (Low-Rank Adaptation) on top of the base model
model = PeftModel.from_pretrained(inference_model, model_path)

Training and evaluation data

kimsan0622/code-knowledge-eval

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.9527	0.9993	178	0.9795	0.5568
0.8867	1.9986	356	0.9173	0.5945
0.7937	2.9979	534	0.9297	0.5982
0.6486	3.9972	712	1.0171	0.5932
0.5555	4.9965	890	1.0441	0.5945

Framework versions

• PEFT 0.11.1
• Transformers 4.44.2
• Pytorch 2.3.0
• Datasets 2.20.0
• Tokenizers 0.19.1

Test set results

Confusion matrix

y_true	pred_0	pred_1	pred_2	pred_3	pred_4	pred_5
0	1112	155	75	35	0	0
1	413	327	274	226	3	1
2	151	207	433	941	37	5
3	71	79	298	3490	928	59
4	14	1	24	1735	3334	1216
5	1	0	2	77	591	1917

Classification reports

y_true	precision	recall	f1-score	support
0	0.63	0.81	0.71	1377
1	0.43	0.26	0.32	1244
2	0.39	0.24	0.30	1774
3	0.54	0.71	0.61	4925
4	0.68	0.53	0.59	6324
5	0.60	0.74	0.66	2588
accuracy			0.58	18232
macro avg	0.54	0.55	0.53	18232
weighted avg	0.58	0.58	0.57	18232

8 bit quantization

Confusion matrix

y_true	pred_0	pred_1	pred_2	pred_3	pred_4	pred_5
0	1107	169	71	29	1	0
1	396	359	273	213	2	1
2	143	234	418	938	36	5
3	66	97	301	3516	888	57
4	12	3	24	1799	3315	1171
5	1	0	2	81	610	1894

Classification reports

y_true	precision	recall	f1-score	support
0	0.64	0.80	0.71	1377
1	0.42	0.29	0.34	1244
2	0.38	0.24	0.29	1774
3	0.53	0.71	0.61	4925
4	0.68	0.52	0.59	6324
5	0.61	0.73	0.66	2588
accuracy			0.58	18232
macro avg	0.54	0.55	0.54	18232
weighted avg	0.58	0.58	0.57	18232