metadata

base_model: unsloth/gpt-oss-20b-unsloth-bnb-4bit
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - gpt_oss
license: apache-2.0
language:
  - en
datasets:
  - microsoft/rStar-Coder

GPT-OSS-Coder-20B

This model is a fine-tuned version of OpenAI's GPT-OSS-20B, optimized for code generation tasks. The fine-tuning leveraged the Unsloth library to enable efficient low-bit quantized training and inference.

Model Details

Base Model: openai/gpt-oss-20b
Training Framework: Hugging Face's TRL library combined with Unsloth optimizations.
Training Data: 1 million randomly generated records, trained for 150 steps

Intended Use

This model is designed to assist with:

Code generation and completion
Programming query answering
Code summarization

About `reasoning_effort`

The reasoning_effort parameter influences the model's focus during text generation:

low: Produces straightforward, concise answers suitable for simple coding tasks.
medium: Balances speed and detail, suitable for moderate complexity tasks.
high: Encourages detailed and complex reasoning, useful for advanced code generation or explanations.

Adjusting this parameter allows you to control the depth of the model's reasoning process, balancing between performance and response complexity.

Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import TextStreamer

tokenizer = AutoTokenizer.from_pretrained("yasserrmd/gpt-oss-coder-20b")
model = AutoModelForCausalLM.from_pretrained("yasserrmd/gpt-oss-coder-20b")

messages = [
    {"role": "system", "content": "You are a helpful coding assistant."},
    {"role": "user", "content": "Using Python to connect MySQL and retrieve table 'employee' where empno is 1234."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    reasoning_effort="low",
).to(model.device)

streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, max_new_tokens=512, streamer=streamer)

Training Overview

The fine-tuning process adapted GPT-OSS-20B to better assist with coding tasks by fine-tuning on a dataset of 1 million random records. The training utilized only the Unsloth library for efficient low-bit quantized fine-tuning.

Citation

@misc{yasserrmd2025gptosscoder20b,
  author = {Yasser RMD},
  title = {GPT-OSS-Coder-20B},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
  url = {https://huggingface.co/yasserrmd/gpt-oss-coder-20b}
}