You can deploy this model and start earning money today!

xap/
Mistral2

Name: Mistral2
Availability: InStock

safetensors

mistral

Views

No views yet

Mistral Instruction Fine-Tuned Model for Code Explanation Semantic Similarity

This repository hosts an instruction fine-tuned large language model for semantic similarity prediction between a student explanation and an expert explanation of a code line.

The model is designed for educational assessment settings where the goal is to measure how closely a student's explanation matches the meaning of an expert reference explanation. Given a code snippet, a student explanation, and an expert explanation, the model returns a similarity score from 0 to 1, where:

0 = minimal or no semantic similarity
1 = maximal semantic similarity

Model Purpose

This model was fine-tuned to support automated evaluation of programming explanations. It is especially useful for tasks such as:

Assessing the quality of student-generated code explanations
Supporting intelligent tutoring systems
Assisting research on automated short-answer scoring
Measuring semantic alignment between novice and expert interpretations of code

Task Definition

The model takes three pieces of information:

A line of code
A student explanation of that code line
An expert explanation of that same code line

It then predicts a single similarity score between 0 and 1.

Input Format

The model expects input in Alpaca-style instruction format.

Dataset Example

json

{
  "instruction": "Assess the semantic similarity of two sentences using a similarity score on a scale from 0 to 1, with 0 indicating minimal or no semantic similarity and 1 representing maximal semantic similarity. Only provide the score without any other output. The input consists of a line of code followed by a student explanation of the line of code and an expert explanation of the line of code as given below.",
  "input": "Code: System.out.println(Duplicate input for number:  + num);\nStudent Explanation: If the two saved adjacent values from the sequence are duplicates, this line will run, printing out that a duplicate of the current number has been found.\nExpert Explanation: This statement prints the duplicate number to the default standard output stream.",
  "output": 0.75,
  "prompt": "Your task is to compute how semantically similar a student explanation of a line of code written in JAVA is compared to an expert explanation of the same line of code. Semantic similarity measures how close the meanings of the two explanations are. The goal is to assess whether the student explains the line of code as well as the expert. ### Instruction: Assess the semantic similarity of two sentences using a similarity score on a scale from 0 to 1, with 0 indicating minimal or no semantic similarity and 1 representing maximal semantic similarity. Only provide the score without any other output. The input consists of a line of code followed by a student explanation of the line of code and an expert explanation of the line of code as given below. ### Input: Code: System.out.println(Duplicate input for number:  + num);\nStudent Explanation: If the two saved adjacent values from the sequence are duplicates, this line will run, printing out that a duplicate of the current number has been found.\nExpert Explanation: This statement prints the duplicate number to the default standard output stream. ### Response:"
}

How to Get Started

Load the model from Hugging Face using transformers:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "xap/Mistral2,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "xap/Mistral2"
)

Inference Example

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "xap/CodeLlama-Instruction-Assessment-new1",
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "xap/CodeLlama-Instruction-Assessment-new1"
)

prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\nFor the given line of code, both the student and expert have provided the explanation for that line of code. Compute the semantic similarity between the student explanation and the expert explanation for the line of code.\n\n### Input:\nFor given line of code int[] values = {5, 8, 4, 78, 95, 12, 1, 0, 6, 35, 46};, the expert explanation is We declare an array of values to hold the numbers. and the student explanation is This line creates the integer array with the values. you need this to achieve the goal bc you need an array to look in\n\n### Response:\n"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=5)

result = tokenizer.decode(outputs, skip_special_tokens=True)
print(result)

Recommended Prompt Template

For best results, format inference inputs as follows:

text

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Assess the semantic similarity of two sentences using a similarity score on a scale from 0 to 1, with 0 indicating minimal or no semantic similarity and 1 representing maximal semantic similarity. Only provide the score without any other output.

### Input:
Code: <code line>
Student Explanation: <student explanation>
Expert Explanation: <expert explanation>

### Response:

Output

The expected output is a single numeric score between 0 and 1.

Example:

0.75

Intended Use

This model is intended for:

Research in educational data mining and NLP
Automated assessment of programming explanations
Semantic similarity scoring in code understanding tasks
Experimental evaluation pipelines for student learning analytics

Limitations

The model is optimized for the fine-tuning format used during training.
Output quality may degrade if prompts differ substantially from the training template.
The model may generate extra text unless prompts explicitly instruct it to return only the score.
The predicted similarity score should be treated as a model estimate, not a definitive ground-truth judgment.

Citation

If this model contributes to your research or application, cite the associated project, paper, or repository as appropriate.

Acknowledgment

This model was developed for semantic assessment of student explanations against expert explanations in programming education contexts.