YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Coding_Model

A lightweight 51M-parameter GPT-2-based language model designed for code generation and experimentation.

The goal of this project is to build a compact coding-focused language model that can be run locally on relatively limited hardware while providing a foundation for further research, fine-tuning, and improvement.

🚧 Early-stage model: This is an experimental model and should not be considered production-ready. More training, evaluation, and improvements are planned.


Model Details

  • Model name: Coding_Model
  • Architecture: GPT-2
  • Parameters: ~51.2M
  • Framework: PyTorch
  • Model format: Safetensors
  • Precision: FP32
  • License: MIT
  • Primary task: Code generation
  • Model type: Causal language model

Why This Model?

Large language models can require significant computational resources to run locally.

Coding_Model explores the idea of building a smaller coding model that can be experimented with and potentially deployed on lower-end hardware.

The goal is not simply to use an existing AI model, but to understand and experiment with the complete model-development process:

Data
  ↓
Tokenization
  ↓
Model Architecture
  ↓
Pretraining
  ↓
Evaluation
  ↓
Fine-tuning
  ↓
Inference
  ↓
Deployment

This model is an early step toward that goal.


Architecture

Coding_Model is based on the GPT-2 architecture.

At a high level:

Input Code
    ↓
Tokenizer
    ↓
Token IDs
    ↓
Token Embeddings
    ↓
Transformer Blocks
    ↓
Self-Attention
    ↓
Feed-Forward Networks
    ↓
Language Modeling Head
    ↓
Next-Token Probabilities

The model performs causal language modeling, meaning it learns to predict the next token based on the tokens that came before it.

For example:

Input:

def add(a, b):
    return

Model:

a + b

The model learns these patterns by training on large amounts of code.


Training Objective

The model is trained using the next-token prediction objective.

Given a sequence:

def add(a, b):

the model attempts to predict subsequent tokens.

During training:

Input tokens
      ↓
Transformer
      ↓
Logits
      ↓
Cross-Entropy Loss
      ↓
Backpropagation
      ↓
Gradient Update

The parameters are updated repeatedly to reduce the training loss.


Training Information

Some training details are still being documented as the project develops.

Configuration Value
Parameters ~51.2M
Architecture GPT-2
Precision FP32
Optimizer [ADD OPTIMIZER]
Learning rate [ADD LEARNING RATE]
Batch size [ADD BATCH SIZE]
Context length [ADD CONTEXT LENGTH]
Training steps [ADD STEPS]
Number of epochs [ADD EPOCHS]
Training tokens [ADD TOKEN COUNT]
Hardware [ADD GPU/HARDWARE]

Dataset

The model is intended for coding-related tasks.

The training data consists of code/text data used to teach the model programming-language syntax, common coding patterns, and relationships between tokens.

Languages

Current/target languages include:

  • Python
  • C++
  • JavaScript
  • Other programming languages depending on the training data

Dataset documentation is still being expanded. Exact dataset sources, filtering procedures, token counts, and preprocessing details will be documented as the project progresses.


Intended Use

Coding_Model is intended for:

  • Learning about language-model training
  • Experimenting with GPT-2 architectures
  • Local code-generation experiments
  • Research and development
  • Fine-tuning experiments
  • Educational purposes
  • Exploring efficient small language models
  • Building lightweight coding assistants

Not Intended For

This model should not currently be relied upon for:

  • Production software development without human review
  • Safety-critical software
  • Autonomous execution of generated code
  • Security-sensitive applications
  • Guaranteed correct programming solutions
  • Generating code without testing

Generated code should always be reviewed, tested, and validated by a developer.


Limitations

Because this is a relatively small and early-stage model, it may:

  • Generate incorrect code
  • Produce incomplete code
  • Repeat patterns
  • Struggle with long programming problems
  • Have limited knowledge of libraries and APIs
  • Generate syntactically incorrect code
  • Hallucinate functions or APIs
  • Perform significantly worse than larger modern coding models

The model's performance depends heavily on the quality and quantity of its training data.


Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "2003Rohansinha/Coding_Model"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = """def factorial(n):
"""

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    temperature=0.7,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Example Prompt

Write a Python function that checks whether a number is prime.

Possible expected style:

def is_prime(n):
    if n < 2:
        return False

    for i in range(2, int(n ** 0.5) + 1):
        if n % i == 0:
            return False

    return True

Actual model output may vary and should always be tested.


Running Locally

The model is designed to be relatively lightweight compared with large language models.

A basic inference workflow is:

User Prompt
    ↓
Tokenizer
    ↓
PyTorch Model
    ↓
GPU / CPU
    ↓
Generated Tokens
    ↓
Decoded Code

The model can be loaded using the Hugging Face Transformers ecosystem.


Evaluation

Formal benchmark results are not yet available.

Future evaluation will measure:

  • Python code generation
  • C++ code generation
  • JavaScript code generation
  • Syntax correctness
  • Compilation success
  • Unit-test pass rate
  • Human evaluation
  • Code completion accuracy
  • Pass@1
  • Pass@k

Planned benchmarks

The project may eventually be evaluated on coding benchmarks such as:

  • HumanEval
  • MBPP
  • Custom coding evaluations

Benchmark results will only be added after reproducible evaluations are performed.


Roadmap

Current

  • Build GPT-2-based architecture
  • Train a ~51M parameter model
  • Upload model to Hugging Face
  • Release model under MIT license

Next

  • Improve training dataset
  • Increase training-token quality
  • Add proper validation dataset
  • Track training/validation loss
  • Evaluate code-generation performance
  • Run HumanEval/MBPP-style benchmarks
  • Improve tokenizer/data pipeline
  • Experiment with learning-rate schedules
  • Experiment with AdamW and other optimizers
  • Improve inference quality
  • Explore quantization
  • Release improved versions

Versioning

The current model should be considered an early experimental release.

Future versions may include:

Coding_Model-v2
Coding_Model-v3
Coding_Model-v4
...

Each major version will aim to improve training data, architecture, evaluation performance, or inference efficiency.


Community

I would love feedback from developers and ML researchers.

Try the model, experiment with it, break it, and let me know where it fails.

Useful feedback includes:

  • Bad generations
  • Good generations
  • Programming languages that work well
  • Programming languages that perform poorly
  • Bugs
  • Suggestions for training data
  • Architecture suggestions
  • Evaluation results

Author

Rohan Sinha

Hugging Face:

2003Rohansinha

Model:

2003Rohansinha/Coding_Model


License

This model is released under the MIT License.

Please review the license and the provenance/licensing of any training data before using the model for commercial applications.


Disclaimer

Coding_Model is an experimental research project.

The generated code may contain bugs, security vulnerabilities, or incorrect assumptions. Always review and test generated code before using it in real applications.


⭐ If You Find This Interesting

This model is only the beginning.

The long-term goal is to explore how far a small, efficient coding model can be pushed through better:

DATA
+
TRAINING
+
ARCHITECTURE
+
EVALUATION
+
OPTIMIZATION

More improvements are coming.

Downloads last month
13
Safetensors
Model size
51.2M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support