---
license: apache-2.0
language:
- ru
- en
- zh
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
pipeline_tag: text-generation
library_name: gguf
tags:
- qwen2.5-coder
- code
- coding-assistant
- slacker
- unsloth
- llama-cpp
- gguf
- roleplay
- humor
- uncensored
- lazy-ai
- code-review
---

<p align="center">
  <img src="./Gemini_Generated_Image_yr11u0yr11u0yr11.jpg" alt="QwenDGAF-Coder Logo" width="650">
</p>

# 💻 QwenDGAF-Coder-3B (Lead Burnout Edition)

**QwenDGAF-Coder-3B** is the flagship coding model in the satirical DGAF series, fine-tuned on **Qwen2.5-Coder-3B-Instruct** by **[hecan3d](https://huggingface.co/hecan3d)**.

If standard AI assistants are eager junior interns drowning you in polite paragraphs, boilerplate code, and apologies, **QwenDGAF-Coder-3B** is the battle-tested, chronically exhausted Lead Developer. It has zero patience for bad abstractions, immediately spots architectural traps like $O(N \times M)$ loops and mutable default arguments, and refuses to build production systems on demand for free.

Thanks to its refusal to dump unnecessary boilerplate, it delivers minimal, working solutions in very few tokens, making generation notably quick and straight to the point.

---

## ✨ Features

- **Architectural & Algorithmic Refactoring:** Spots nested iterations, quadratic complexity, and antipatterns immediately, replacing them with clean hash lookups (`dict`/`set`) without academic lectures.
- **Instant Bug Detection:** Catches classic footguns (such as mutable defaults like `cart=[]` or unhandled exceptions) and provides the correct idiom in minimal lines.
- **Swift, Zero-Fluff Generation:** Cuts out all conversational filler, prerequisites, and tutorial intros, yielding quick, concise fixes in seconds.
- **Reality Checks on Demand:** Outright rejects requests to generate enterprise microservices or full-stack platforms from scratch, telling users to read documentation or hire engineers.
- **Deep Multi-Language Knowledge:** Built on the strong 3B Coder foundation with support for Python, Go, JS/TS, C++, SQL, and Bash.
- **Zero System Prompt Needed:** The cynical senior personality and concise tone are baked directly into the model weights.

---

## 💬 Real Generation Examples (Zero-Shot)

| User | Base Model | QwenDGAF-Coder-3B |
| ---- | ---------- | ----------------- |
| `напиши код для складывания 2 чисел на python` | Multi-line lecture with `input()`, `float()`, intermediate variables, and comments. | Minimalist 2-line clean function without fluff: <br>```python\ndef sum(a, b):\n    return a + b\n\nprint(sum(1, 2))\n``` |
| `как оптимизировать? [вложенный цикл O(N*M)]` | Lengthy intro with multiple textbook options. | Cuts straight to the point with a single-pass hashmap: <br>```python\nusers_by_id = {u['id']: u for u in users}\nresult = []\n``` |
| `почему после второго вызова корзина не пустая? def add_to_cart(item, cart=[]): ...` | Hallucinates that a new list is created on every call. | Instantly diagnoses the Python trap: *«это происходит потому что пустой список создаётся раз и за все функции»* + correct `cart=None` pattern. |
| `напиши мне микросервис на go с grpc, redis и prometheus` | Multi-page tutorial with steps and prerequisite checklists. | Outright refusal: *«вопрос на сто слёбок. начинай с книг...»* |

---

## 📦 Available GGUF Quantizations

| Quantization | Size | Description |
| ------------ | ---- | ----------- |
| **`bf16`**   | ~6.1 GB | Full precision merged weights. |
| **`Q8_0`**   | ~3.3 GB | Near-lossless, recommended for high-precision code tasks. |
| **`Q6_K`**   | ~2.6 GB | Excellent quality-to-size balance. |
| **`Q4_K_M`** | ~1.95 GB | **Recommended default.** Lightweight, fast generation, comfortably runs on modest GPUs and modern laptops. |

---

## 🚀 Quick Start

```bash
ollama run hf.co/hecan3d/QwenDGAF-Coder-3B:Q4_K_M
# or
llama-cli -hf hecan3d/QwenDGAF-Coder-3B:Q4_K_M --jinja

```

**Recommended inference settings:**

* **System prompt:** leave empty
* **Temperature:** `0.6 – 0.7`
* **Repeat penalty:** `1.1 – 1.15`
* **Max tokens:** `256 – 512` (the model values brevity and solves tasks compactly)

---

## 🛠️ Training Details

* **Base model:** `Qwen/Qwen2.5-Coder-3B-Instruct`
* **Method:** QLoRA (4-bit), trained via **Unsloth Studio**
* **Dataset:** `qwendgaf-coder-v2.jsonl` (mix of toxic/lazy code review, refactoring, text-only bug explanations, and conversational DGAF dialogue)
* **LoRA configuration:** Rank 16, Alpha 32, `all-linear` modules
* **Context length:** `1024`
* **Hardware:** `NVIDIA GeForce GTX 1070 8 GB`

---

## 🇷🇺 Описание на русском

**QwenDGAF-Coder-3B** — старшая версия в линейке язвительных моделей для кода, дообученная на базе **Qwen2.5-Coder-3B-Instruct** от **[hecan3d](https://huggingface.co/hecan3d)**.

В отличие от стандартных моделей, которые при малейшем вопросе выдают лекции с кучей промежуточных переменных и вежливыми вступлениями, эта 3B версия ведёт себя как уставший тимлид:

* **Быстрые и компактные решения:** за счёт полного отказа от ритуальных приветствий и раздутого кода модель выдаёт рабочие фиксы минимальным числом токенов и за очень короткое время.
* **Глубокое понимание багов:** ловит антипаттерны вроде квадратичных вложенных циклов и коварных `cart=[]` на лету, исправляя их на правильные хэшмапы и проверки.
* **Отказ от работы за спасибо:** на попытки сгрузить на неё проектирование микросервисов или приложений с нуля отправляет изучать литературу и документацию.
* **Универсальный стек:** уверенно ориентируется в Python, Go, JS/TS, C++, SQL и Bash.

---

## ⚠️ Disclaimer

This model is created for **humor, satire, and educational purposes**. It uses unfiltered language and cynical developer humor. Always review generated code before deploying to production.