metadata

datasets:
  - nvidia/OpenCodeReasoning-2
  - GetSoloTech/Code-Reasoning
base_model: GetSoloTech/GPT-OSS-Code-Reasoning-20B
library_name: mlx
tags:
  - code-reasoning
  - coding
  - reasoning
  - problem-solving
  - algorithms
  - python
  - c++
  - competitive-programming
  - vllm
  - mlx
pipeline_tag: text-generation

GPT-OSS-Code-Reasoning-20B-qx86-hi-mlx

This is an experimental quant with mixed precision selective layers, some compressed to 6bit, all rendered with group size 32

Side effects of quanting with the qx86-hi formula

I needed Haskell code.

The q6 starts with Haskell, 10k tokens down the road writes Python, and finishes with React at 30k tokens
The q6-hi, encoded with group size 32, writes some Haskell, and stops around 24k tokens
The qx86-hi worked for 43k tokens, reasoning around the Haskell solution without skipping a beat

It didn't forget Python, it's just a bit more open-minded about other languages

From the original model card:

Overview

Base model: openai/gpt-oss-20b
Objective: Supervised fine-tuning for competitive programming and algorithmic reasoning
Dataset: nvidia/OpenCodeReasoning-2 (OCR-2), combining python and cpp splits. 
Each sample reconstructs the upstream question and uses the dataset's r1_generation as the assistant response
Context length: 4096 tokens
Training method: LoRA SFT via TRL SFTTrainer

Intended Use

Intended: Generating Python/C++ solutions and reasoning for competitive programming tasks
Out of scope: Safety-critical applications. May hallucinate or produce incorrect/inefficient code

This model GPT-OSS-Code-Reasoning-20B-qx86-hi-mlx was converted to MLX format from GetSoloTech/GPT-OSS-Code-Reasoning-20B using mlx-lm version 0.26.4.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("GPT-OSS-Code-Reasoning-20B-qx86-hi-mlx")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)