GPT OSS LOOP 20B

GPT OSS 20B with the LOOP V2 behavioral prompt layer for structured self-checking, requirement tracking, and answer refinement.

This is the more accessible companion to the 120B version, intended for users with less powerful hardware who still want to experiment with the LOOP approach.

What this is

This model uses GPT OSS 20B with an added system-prompt layer called LOOP V2.

LOOP V2 encourages the model to:

• Identify the user’s actual intent and practical goal
• Track requirements, constraints, and requested formats
• Distinguish evidence from assumptions and unknowns
• Check whether the available information is sufficient
• Review responses for omissions, contradictions, and unsupported claims
• Correct issues before presenting the final answer
• Return a finished answer without narrating the internal quality process

No additional fine-tuning was performed as part of this publication.

Usage

ollama pull optimistic_mcnulty_933/gpt-oss-loop-20b:latest
ollama run optimistic_mcnulty_933/gpt-oss-loop-20b:latest

The model can also be used through the Ollama API:

{
  "model": "optimistic_mcnulty_933/gpt-oss-loop-20b:latest",
  "messages": [
    {
      "role": "user",
      "content": "Your question here"
    }
  ],
  "stream": false
}

Intended use

This model is suitable for general assistance, planning, writing, coding, brainstorming, and multi-step tasks where organization and requirement tracking are useful.

The 20B version is easier to run than the 120B version, but it should not be expected to produce equivalent results on every difficult reasoning task.

Important limitations

• LOOP is a behavioral prompt layer, not a guarantee of accuracy.
• The model can still make factual, reasoning, or calculation errors.
• Self-generated claims about the model’s own identity or training should not be treated as authoritative. Use the model page and published tag as the source of truth.
• Current, specialized, legal, medical, financial, and safety-critical information should be independently verified.
• Results depend on available context, prompt quality, Ollama configuration, and hardware.
• This is an independent community publication and is not an official OpenAI model release.

Feedback

This is an experimental public model intended to be tested and improved. If you find inconsistent behavior, missed requirements, or useful improvements to the LOOP instructions, feedback is welcome.

The goal is to make structured and careful model behavior easier for others to test, adapt, and improve.