---
language:
  - en
  - code
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen2.5-Coder-0.5B-Base
tags:
  - code
  - python
  - code-generation
  - text-generation
  - qwen2.5
  - qwen
  - llama.cpp
  - gguf
  - unsloth
  - qlora
  - lora
  - small-models
  - edge-ai
  - local-ai
  - minsore
  - ukrainian
  - experimental
  - 0.5b
  - generation
model-index:
  - name: Quill Gen 1
    results:
      - task:
          type: text-generation
          name: Code Generation
        dataset:
          type: openai_humaneval
          name: HumanEval@50
        metrics:
          - type: pass@1
            value: 80.0
            name: Pass@1
---

# 🟢 Quill Gen 1

Lightweight Python code generation model built on Qwen2.5-Coder-0.5B.

Quill Gen 1 is a 0.5B parameter experimental model for generating Python functions from natural-language descriptions. It is the younger sibling of Pepper 1 Preview — same task, one third the size. Trained as an experiment in small-model code generation, it reaches 80% on HumanEval@50 while running comfortably on consumer hardware.

> ⚠️ **This is NOT a FIM model.** Quill Gen 1 cannot do fill-in-the-middle completion. For autocomplete use [Quill 1 Preview](https://huggingface.co/Minsore/quill-1-preview).

    Part of the Minsore family · minsore.com

## ✨ Highlights

- 🧠 **Instruction-tuned** — writes Python functions from plain descriptions
- 📦 **Tiny** — 0.5B params, ~400 MB in Q4_K_M
- 🎯 **Purpose-built** — code generation only, not chat, not FIM
- ⚡ **Fast** — designed for low-VRAM setups
- 🧪 **Experimental** — a proof of concept for small-model generation
- 🆓 Apache 2.0 — same license as base model

## 📊 Benchmarks

![Quill Gen comparison](assets/quill_gen_1_benchmarks.png)

Evaluated against 0.5B–1.5B code models. All runs used `temperature=0.0`, `max_tokens=512`, `--chat-template none`.

| Benchmark | Quill Gen 1 | Pepper 1 Preview | Quill 1 Preview | Qwen2.5-Coder 0.5B |
|---|---|---|---|---|
| HumanEval@50 | **80.0%** | 82.0% | 54.0% | 52.0% |
| HumanEval-Infilling (EditSim) | 0.161 | — | **0.423** | 0.045 |
| HumanEval-Infilling (Exact Match) | 0% | — | **16%** | 0% |
| Delulu FIM (EditSim) | 0.048 | — | **0.318** | 0.035 |

> 📌 **Key insight:** Quill Gen 1 reaches **80% HumanEval@50** — nearly matching Pepper 1 Preview (82%) at **one third the parameter count**. On FIM-specific tasks it underperforms significantly, which is expected: this model was trained for generation, not autocomplete.

## 🚀 Quick Start

### llama.cpp

```bash
llama-server -m quill-gen-1.Q4_K_M.gguf \
  --port 8080 \
  -ngl 99 \
  -c 768 \
  --chat-template none
```

**Requirements:** any GPU with ≥2 GB VRAM (full offload), or partial CPU offload as fallback.

### Request

```bash
curl http://localhost:8080/completion \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "### Instruction:\nWrite a Python function that checks if a number is prime.\n\n### Response:\n",
    "n_predict": 256,
    "temperature": 0.0
  }'
```

### Python

```python
import requests

def generate(instruction, max_tokens=256):
    prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"
    r = requests.post("http://localhost:8080/completion", json={
        "prompt": prompt,
        "n_predict": max_tokens,
        "temperature": 0.0,
        "stop": ["### Instruction:", "<|endoftext|>"],
    })
    return r.json()["content"]

print(generate("Write a Python function that reverses a string."))
# → def reverse_string(s):
#       return s[::-1]
```

## ⚙️ Recommended Settings

| Parameter | Value | Notes |
|---|---|---|
| `--chat-template` | `none` | Required. Quill Gen 1 is not a chat model |
| `n_predict` | 128–256 | Works best on short completions |
| `temperature` | `0.0` | Deterministic; use `0.2` for variation |
| `repeat_penalty` | `1.1` | Prevents repetition |
| `-c` | `768` | Matches training context length |

## ⚠️ Limitations

- **Generation only** — does not support fill-in-the-middle, tool calling, or chat
- **Python-only** — trained exclusively on Python code
- **Small context** — 768 tokens; long files are truncated
- **Weak on FIM** — EditSim 0.16, Exact Match 0% (use Quill 1 instead)
- **Occasional over-explanation** — may include comments when only code is requested
- **Number looping** — may repeat large integers on some prompts (use `repeat_penalty=1.1`)

## 🧬 Training Details

| | |
|---|---|
| Base model | Qwen2.5-Coder-0.5B-Base |
| Method | QLoRA (r=16, α=16) |
| Data | FIM-converted Python instruction data |
| Epochs | 1 |
| Context | 768 tokens |
| Optimizer | `paged_adamw_8bit` |

## 📁 Files

| File | Size | Description |
|---|---|---|
| `quill-gen-1.Q4_K_M.gguf` | ~400 MB | Ready to use with llama.cpp (recommended) |
| `model.safetensors` | ~1 GB | Full-precision merged weights |
| `config.json` | — | Model config |
| `tokenizer.json` | — | Tokenizer |
| `tokenizer_config.json` | — | Tokenizer config |
| `generation_config.json` | — | Generation defaults (optional) |

## 🗺️ Roadmap

- **Quill 2** — FIM autocomplete, fixed suffix handling, JS/TS/Rust support
- **Pepper 2** — improved MBPP and LiveCodeBench
- **Symphony** — flagship agentic code model (3B MoE)

## 📜 License

Apache 2.0 — same as the base Qwen2.5-Coder-0.5B model.

## 🙏 Credits

- Base model: [Qwen2.5-Coder-0.5B-Base](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Base) by Alibaba Cloud
- Training framework: Unsloth
- Inference: llama.cpp

## 📬 Contact

Minsore — Ukrainian AI lab building open language models.

- 🌐 [minsore.com](https://minsore.com)
- 🤗 [huggingface.co/Minsore](https://huggingface.co/Minsore)
- 💬 Built by [@Sollamon](https://huggingface.co/Sollamon)

⭐ If Quill Gen 1 is useful, star the repo and share your results.