README.md

---
language:
  - en
  - code
license: apache-2.0
library_name: llama.cpp
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
tags:
  - code
  - code-assistant
  - instruction-following
  - python
  - qwen2.5-coder
  - minsore
  - gguf
pipeline_tag: text-generation
---

# 🌶️ Pepper 1 Preview

**Lightweight instruction-following code assistant for Python, built on Qwen2.5-Coder-1.5B.**

Pepper is a 1.5B parameter code assistant designed for writing functions from natural-language descriptions. It runs on consumer hardware and is one of the few publicly available 1.5B models tuned specifically for instruction-following code generation.

> Part of the **Minsore** family · [minsore.com](https://minsore.com)

---

## ✨ Highlights

- 🧠 **Instruction-tuned** — writes functions from plain English descriptions
- 🐍 **Python-focused** — clean, idiomatic code without fluff
- ⚡ **Real-time** — designed for interactive use
- 📦 **Compact** — 1.5B params, ~1 GB in Q4_K_M
- 🎯 **Purpose-built** — trained for code generation, not chat
- 🆓 **Apache 2.0** — same license as base model

---

![Pepper comparison](assets/pepper_comparison.png)

## 📊 Benchmarks

Evaluated against 1–1.5B code models. All runs used `temperature=0.0`, `max_tokens=512`, `--chat-template none`.

| Benchmark | Pepper 1 Preview | Base Qwen 1.5B | Llama-3.2-1B-Code | Yi-Coder-1.5B |
|---|---:|---:|---:|---:|
| **HumanEval@50** | **82.0%** | 75.0% | 64.0% | 24.0% |
| **MBPP@50** | 26.0% | 20.0% | 16.0% | 38.0% |
| **LiveCodeBench@30** | 20.0% | 30.0% | 10.0% | 26.7% |
| **BigCodeBench@30** | 23.3% | 23.3% | 13.3% | 30.0% |

> 📌 **Key insight:** Pepper 1 Preview outperforms the base model on **HumanEval by +7 points** — the primary benchmark for instruction-following code generation. On MBPP, both models underperform the official numbers because this evaluation uses zero-shot prompting with `--chat-template none`. Yi-Coder tested without its native chat template — its HumanEval score reflects format mismatch, not model quality.

---

## 🚀 Quick Start

### llama.cpp

```bash
llama-server -m pepper-1-preview.Q4_K_M.gguf \
  --port 8080 \
  -ngl 99 \
  -c 4096 \
  --chat-template chatml
```

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

### Chat request

```bash
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "Write a Python function that checks if a number is prime."}
    ],
    "max_tokens": 512,
    "temperature": 0.0
  }'
```

### Python

```python
import requests

def ask(prompt, max_tokens=512):
    r = requests.post("http://localhost:8080/v1/chat/completions", json={
        "messages": [{"role": "user", "content": prompt}],
        "max_tokens": max_tokens,
        "temperature": 0.0,
    })
    return r.json()["choices"][0]["message"]["content"]

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

---

## ⚙️ Recommended Settings

| Parameter | Value | Notes |
|---|---|---|
| `--chat-template` | `chatml` | **Required.** Pepper expects ChatML. |
| `n_predict` | 256–1024 | 512 is a good default |
| `temperature` | 0.0 | Deterministic; use 0.2 for variation |
| `repeat_penalty` | 1.1 | Prevents repetition |
| `-c` | 4096 | Longer context available but degrades |

---

## ⚠️ Limitations

- **Python-only** — not trained on other languages
- **No FIM support** — for fill-in-the-middle tasks, use [Quill](https://huggingface.co/minsore/quill-1-preview)
- **Not an agent** — does not support tool calling or multi-step planning
- **Weak on algorithmic tasks** — LiveCodeBench score below base Qwen
- **4K inference context** — long files are truncated
- **Occasional over-explanation** — may include comments when only code is requested

---

## 🧬 Training Details

| | |
|---|---|
| **Base model** | Qwen2.5-Coder-1.5B-Instruct |
| **Method** | QLoRA fine-tuning on curated Python code data |
| **Context** | 4096 tokens (inference) |

---

## 📁 Files

| File | Size | Description |
|---|---:|---|
| `pepper-1-preview.Q4_K_M.gguf` | ~1 GB | Ready to use with `llama.cpp` |
| `model.safetensors` | ~3 GB | Full precision (transformers) |
| `config.json`, `tokenizer.json` | — | Config for transformers |

---

## 🗺️ Roadmap

- **Pepper 2** — improve MBPP and LiveCodeBench, add multi-language support
- **Quill 2** — extend FIM to JS/TS/Rust
- **Symphony** — flagship code model (3B)

---

## 📜 License

Apache 2.0 — same as the base Qwen2.5-Coder-1.5B model.

---

## 🙏 Credits

- Base model: [Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct) by Alibaba Cloud
- Training framework: [Unsloth](https://github.com/unslothai/unsloth)
- Inference: [llama.cpp](https://github.com/ggerganov/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)

---

<div align="center">

**⭐ If Pepper is useful, star the repo and share your results.**

</div>