codex-finetune

This model is a fine-tuned version of bigcode/starcoder2-3b on the code_search_net dataset.

Model description

It’s designed to serve as an intelligent coding copilot: generate code, explain functions, refactor logic, and complete partial implementations.

🚀 Features

• Multi-task formatting: Instruction-tuned samples with tasks like code generation, docstring generation, function completion, and code improvement.
• Efficient LoRA training using PEFT and transformers.
• Token-level preprocessing with Hugging Face's tokenizer and trainer utilities. Training tracked via Weights & Biases (W&B).
• Dataset sampling + tokenization to stay memory-efficient.
• Ready for inference integration and API deployment.

Dataset

• Source: code_search_net (Python split)
• Fields Used: func_code_string, func_documentation_string
• Size after sampling:
  • Train: 1000 samples
  • Validation: 200 samples
  • Test: 200 samples

Format: Multi-Task Examples

Examples were formatted into prompts like:

Instruction: Write a function for this description: "Calculate factorial recursively."

Response: def factorial(n): return 1 if n == 0 else n * factorial(n - 1)

Model

• Base: bigcode/starcoder2-3b

• PEFT Config:

  • r=8, lora_alpha=16
  • target_modules=["q_proj", "v_proj"]
  • dropout=0.05, bias="none"

• Training Config:

  • per_device_train_batch_size=4
  • num_train_epochs=3
  • learning_rate=2e-4
  • save_steps=100
  • logging_dir=./logs

Dependencies

pip install transformers peft datasets accelerate wandb
pip install bitsandbytes


### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- num_epochs: 3
- mixed_precision_training: Native AMP

### Training results
Step	Training Loss
500	1.700700
1000	1.305100
1500	1.234500
 [3000/3000 1:11:12, Epoch 3/3]
Step	Training Loss
500	1.700700
1000	1.305100
1500	1.234500
2000	1.229400
2500	1.185200
3000	1.203400


### Framework versions

- PEFT 0.15.2
- Transformers 4.52.4
- Pytorch 2.6.0+cu124
- Datasets 3.6.0
- Tokenizers 0.21.1

###  Run Locally

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("khushimalik53/codex-finetune")
model = AutoModelForCausalLM.from_pretrained("khushimalik53/codex-finetune")

prompt = "### Instruction:\nExplain what this function does:\ndef reverse_string(s): return s[::-1]\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(output[0], skip_special_tokens=True))