---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - code
  - code-review
  - programming
  - qlora
  - unsloth
  - qwen2.5
  - bug-detection
datasets:
  - sahil2801/CodeAlpaca-20k
language:
  - en
pipeline_tag: text-generation
library_name: transformers
model-index:
  - name: CodeLens-7B
    results: []
---

# CodeLens-7B

A fine-tuned **Qwen2.5-7B-Instruct** model specialized for **code review, bug detection, and programming assistance**. It analyzes code snippets, identifies issues, suggests improvements, and writes clean solutions across multiple programming languages.

## Key Details

| | |
|---|---|
| **Base model** | Qwen/Qwen2.5-7B-Instruct |
| **Method** | QLoRA (4-bit NF4, rank 16, alpha 16) |
| **Library** | Unsloth + TRL SFTTrainer |
| **Dataset** | sahil2801/CodeAlpaca-20k (10K examples) |
| **Hardware** | NVIDIA RTX A5000 (24GB VRAM) on RunPod |
| **Training time** | ~2.65 hours (500 steps) |
| **Final loss** | 0.450 |
| **Parameters trained** | 40.4M of 7.66B (0.53%) |
| **Format** | ChatML |
| **Output** | Merged 16-bit safetensors |

## Dataset

Trained on 10,000 examples from [sahil2801/CodeAlpaca-20k](https://huggingface.co/datasets/sahil2801/CodeAlpaca-20k), a code instruction-following dataset covering code generation, debugging, explanation, and review tasks across Python, JavaScript, Java, C, SQL, and more.

## Usage

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("sriksven/CodeLens-7B")
tokenizer = AutoTokenizer.from_pretrained("sriksven/CodeLens-7B")

messages = [
    {
        "role": "system",
        "content": "You are an expert code reviewer and programmer. Analyze code, find bugs, suggest improvements, and write clean efficient solutions.",
    },
    {
        "role": "user",
        "content": "Review this Python function for bugs and improvements:\n\ndef find_duplicates(lst):\n    seen = []\n    dupes = []\n    for i in lst:\n        if i in seen:\n            dupes.append(i)\n        seen.append(i)\n    return dupes",
    },
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Unsloth (faster inference)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="sriksven/CodeLens-7B",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
```

## Capabilities

- **Code review** — analyze code for bugs, anti-patterns, and style issues
- **Bug detection** — identify logical errors, off-by-one mistakes, edge cases
- **Code generation** — write functions, classes, and scripts from descriptions
- **Code explanation** — explain what a piece of code does step by step
- **Refactoring suggestions** — propose cleaner, more efficient alternatives
- **Multi-language** — Python, JavaScript, Java, C/C++, SQL, HTML/CSS, and more

## Intended Use

- Local code review assistant
- Programming tutoring and education
- Code quality tooling in CI/CD pipelines
- Prototyping developer tools with local LLMs

## Limitations

- Trained on instruction-following code data, not real code review conversations from PRs
- May not catch security vulnerabilities that require deep context
- Code suggestions should be tested before use in production
- Best with shorter code snippets (functions/classes) rather than full files
- No execution or testing capability — suggestions are pattern-based

## Training Metrics

Loss decreased steadily from 2.17 to 0.27 over 500 steps (~13 epochs), indicating strong learning on the code instruction data.

| Step | Loss | Epoch |
|---|---|---|
| 10 | 2.168 | 0.26 |
| 100 | 0.503 | 2.05 |
| 250 | 0.430 | 6.41 |
| 400 | 0.310 | 10.26 |
| 500 | 0.278 | 12.83 |

## Training Infrastructure

| | |
|---|---|
| **GPU** | NVIDIA RTX A5000 24GB |
| **Cloud** | RunPod ($0.27/hr) |
| **Framework** | Unsloth 2026.5.2 + TRL + Transformers 5.5.0 |
| **Precision** | BF16 training, 4-bit NF4 base quantization |
| **Optimizer** | AdamW 8-bit |
| **Learning rate** | 2e-4, linear decay |
| **Batch size** | 16 effective (4 per device × 4 accumulation) |
| **Packing** | Enabled |

## Source Code

Training scripts: [github.com/sriksven/LLM-FineTune-Suite](https://github.com/sriksven/LLM-FineTune-Suite)

## License

Apache 2.0