🔥 Pyroton

A lightweight Python code generation model fine-tuned from Qwen2.5-Coder-0.5B-Instruct.

---

Overview

Pyroton is a lightweight Python-focused code generation model fine-tuned from Qwen/Qwen2.5-Coder-0.5B-Instruct using supervised fine-tuning (SFT) on Python instruction-style datasets.

The goal is to create a small, efficient model that handles easy to medium Python tasks while remaining practical for free-tier GPUs and lightweight deployment including mobile phones.

---

Model Variants

Base adapter

• shohuu/pyroton — general Python instruction-tuned adapter

Prime-fix patched adapter

• shohuu/pyroton-primefix-v3 — targeted repair finetuning for correctness bugs (recommended)

---

Quick Start

Load latest patched adapter (recommended)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-0.5B-Instruct",
    dtype=torch.bfloat16,
    device_map="auto",
)

model = PeftModel.from_pretrained(base, "shohuu/pyroton-primefix-v3")
tokenizer = AutoTokenizer.from_pretrained("shohuu/pyroton-primefix-v3")
tokenizer.pad_token = tokenizer.eos_token

prompt = "### Instruction:\nWrite a Python function to reverse a string\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=220,
    do_sample=False,
    repetition_penalty=1.1,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Run GGUF locally (Ollama / LM Studio / PocketPal)

# Download pyroton-q4.gguf from Files tab
# Then load in your preferred local LLM app

---

Recommended Inference Settings

Correctness mode (recommended)

Most reliable for code generation:

outputs = model.generate(
    **inputs,
    max_new_tokens=220,
    do_sample=False,
    repetition_penalty=1.1,
)

Sampled mode (more variety, less stable)

outputs = model.generate(
    **inputs,
    max_new_tokens=220,
    do_sample=True,
    temperature=0.1,
    top_p=0.9,
    repetition_penalty=1.2,
)

---

Example Output

Prompt:

Write a Python function to check if a number is prime

Pyroton Output:

import math

def is_prime(n):
    """Check if the given integer n is prime."""
    if n <= 1:
        return False
    for i in range(2, int(math.sqrt(n)) + 1):
        if n % i == 0:
            return False
    return True

---

Training Details

Setting	Value
Base Model	Qwen2.5-Coder-0.5B-Instruct
Datasets	python_code_instructions_18k_alpaca, CodeAlpaca-20k, code_instructions_122k_alpaca_style
Total Samples	~95,362 (Python-filtered)
Training Strategy	Chunked SFT (5 chunks)
LoRA Rank	16
LoRA Alpha	32
Batch Size	2
Gradient Accumulation	8
Learning Rate	1e-4
Precision	BFloat16
Max Length	512
Final Training Loss	~0.712

Repair finetuning

After main training, targeted repair finetuning was applied to fix:

• Missing import math / math.sqrt issues
• Incorrect handling of edge cases (negative numbers, 0, 1)
• Latest patched adapter: shohuu/pyroton-primefix-v3

Evaluation

Tested against execution-based harness on is_prime() with inputs: -1, 0, 1, 2, 3, 4, 6, 9, 17, 49

• Greedy decoding: 5/5 passing ✅
• Sampled decoding: improved but less stable

---

GGUF / Mobile Deployment

Pyroton is available as a GGUF file for local deployment:

File	Quantization	Size
pyroton-q4.gguf	Q4_K_M	~397MB

Compatible apps:

• PocketPal AI (Android/iOS) — search shohuu/Pyroton
• LM Studio (Desktop)
• Ollama (Desktop)

---

Known Limitations

• 0.5B model — can degrade on harder tasks or complex reasoning
• Greedy decoding is more reliable than sampling for correctness
• Most thoroughly tested on short Python coding tasks
• Broader evaluation across more libraries still needed

---

GitHub

github.com/TunasTuna/pyroton

---

Requirements

transformers
datasets
trl
peft
bitsandbytes
accelerate
torchao

---

License

Apache 2.0 — see LICENSE for details.

Base model (Qwen2.5-Coder) is also Apache 2.0. Attribution to Alibaba Cloud / Qwen Team.

---

Acknowledgements

• Qwen Team for the base model
• iamtarun for the original dataset
• Hugging Face for the training ecosystem
• My friend Yumi for the name 🔥