---
title: Linus Coder
language:
  - en
license: mit
library_name: transformers
tags:
  - text-generation
  - lora
  - qwen2.5-coder
  - code-review
  - kernel-patching
datasets:
  - linus-torvalds-reviews
base_model: unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit
---

# Linus Coder

A Qwen2.5-Coder-7B model fine-tuned with LoRA to review Linux kernel patches in the blunt, deeply technical style of Linus Torvalds.

## Intended Use

Feed a patch diff and get a Linus-style review ending with `Verdict: <approve|request_changes|reject|question>`.

## Input format

```
Reviewer style target: linus-torvalds
Subsystem: <subsystem>
Patch subject: <subject>

<patch diff>

Write your review of this patch.
End your review with exactly: Verdict: <approve|request_changes|reject|question>
```

## Usage (merged model)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("akshatkot/linus-coder")
model = AutoModelForCausalLM.from_pretrained("akshatkot/linus-coder", torch_dtype=torch.float16)
```

## Training

Fine-tuned with Unsloth SFTTrainer, 2 epochs, lr=1e-4, cosine schedule on LKML patch reviews.

## License

MIT