EpistemeAI/metatune-gpt20b-R1.1

EpistemeAI/metatune-gpt20b-R1.1: A Self-Improving LLM

EpistemeAI's metatune-gpt20b-R1.1 is a 20 billion parameter language model, fine-tuned from OpenAI's gpt-oss-20b, designed for general-purpose applications. Its primary differentiator is its recursive self-improvement mechanism, allowing it to generate new training data, evaluate its own performance, and adapt its hyperparameters to enhance capabilities. This release is the 5th metacycle (generation) of this continuous improvement process.

Key Capabilities

• Recursive Self-Improvement: Continuously learns and optimizes itself without external human intervention in the training loop.
• General Purpose: Suitable for a broad spectrum of natural language processing tasks.
• Adjustable Reasoning Levels: Users can set reasoning levels (Low, Medium, High) via system prompts to balance response speed and detail, with 'High' recommended for robust performance and guardrails.
• Tool Use: Excels in agentic operations, including web browsing with built-in tools and function calling with defined schemas.
• Fine-tuning: The base gpt-oss models are amenable to further fine-tuning for specialized use cases.

Performance Highlights

While specific comprehensive benchmarks are not provided, the model shows iterative improvements across generations. For instance, on gsm8k_cot_llama, metatune R1.1 achieved +1.0 exact match, compared to 0.9796 for metatune R1 and 0.91 for metatune R0 (on a limited 10-shot evaluation).

Guardrails and Safety

Users are advised to set reasoning="high" in prompts to mitigate jailbreaking and prompt injection. For enhanced safety, EpistemeAI recommends using openai/gpt-oss-safeguard-20b as a preliminary guardrail model.