viperdevAI's picture
Viperfish Sunlight 1.5B (MLX 4-bit)
d8218dc verified
|
Raw History Blame Contribute Delete
2.88 kB
metadata
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
library_name: mlx
pipeline_tag: text-generation
language:
  - en
tags:
  - mlx
  - swift
  - swiftui
  - code
  - lora

Viperfish Sunlight 1.5B

Sunlight is the smallest model in the Viperfish family of open-weight models. It is a Swift / SwiftUI-focused fine-tune of Qwen2.5-Coder-1.5B-Instruct, small enough to run on an 8 GB Mac.

This repo is the MLX 4-bit build (about 840 MB): the LoRA adapter fused into the 4-bit base. It runs with mlx-lm on Apple Silicon.

Quick start

pip install mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("viperdevAI/Viperfish-Sunlight-1.5B")
messages = [{"role": "user", "content": "Write a SwiftUI view with a counter and an increment button."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

Details

  • Base model: Qwen/Qwen2.5-Coder-1.5B-Instruct (Apache-2.0), 4-bit MLX quantization. All credit for the base model goes to the Qwen team.
  • Method: LoRA (rank 8, scale 20, 16 layers), 6,164 steps at batch size 1, sequence length 1024, learning rate 5e-5 with cosine decay.
  • Training data: 6,164 Swift / SwiftUI source files from The Stack v1 (dedup). The dataset is not redistributed here; upstream licenses apply to the original files.
  • License: Apache-2.0.

Results

Compared with the base model on my own benchmarks. Swift answers are scored by whether the generated code compiles with swiftc; HumanEval is the standard Python benchmark.

Model SwiftUI (60 prompts) Plain Swift (33 prompts) HumanEval (164)
Qwen2.5-Coder-1.5B-Instruct (4-bit) 16 13 67.7% (111)
Viperfish Sunlight 1.5B 31 14 56.1% (92)

SwiftUI roughly doubles. Plain Swift barely moves. Python gets worse.

Limitations

  • Python regression: HumanEval drops from 67.7% to 56.1%. If you want a general-purpose coding model, use the base model; Sunlight trades general ability for SwiftUI.
  • Training data was raw source code, not chat. The model often stops without closing its final code fence, and long answers may run on.
  • It is a 1.5B model: expect mistakes, invented APIs and outdated SwiftUI idioms. Always compile and review generated code.
  • Trained only on Swift / SwiftUI; it may not generalize to other languages.
  • The benchmark prompt sets are my own and small; treat the numbers as indicative, not definitive.

The Viperfish family

Sunlight 1.5B, Twilight 3B, Midnight 7B and Hadal 14B, named after the ocean's depth zones. More at viperfish.dev.