EpistemeAI/Codeforce-metatune-gpt20b
EpistemeAI/Codeforce-metatune-gpt20b is a 20 billion parameter language model, fine-tuned from OpenAI's gpt-oss-20b, featuring a 32768 token context length. This model is notable for being an early public release of a recursive self-improving AI, capable of generating its own Codeforce-Cot data, evaluating performance, and adjusting hyperparameters. It excels in coding tasks, demonstrating superior reasoning and code synthesis capabilities.
Loading preview...
Model Overview
EpistemeAI/Codeforce-metatune-gpt20b is a 20 billion parameter model, fine-tuned from OpenAI's gpt-oss-20b, and represents an early public release of a recursive self-improving AI. It is designed to enhance its own performance by generating new data, evaluating its output, and adaptively adjusting its hyperparameters. This model is particularly optimized for coding tasks and demonstrates strong reasoning abilities.
Key Capabilities & Features
- Recursive Self-Improvement: Generates its own Codeforce-Cot data, evaluates performance, and adjusts hyperparameters for continuous improvement.
- Coding Excellence: Achieves a HumanEval score of 90, outperforming larger models like Qwen 3 235B and DeepSeek-R1 70B in coding problem-solving and reasoning.
- Adjustable Reasoning Levels: Supports 'Low', 'Medium', and 'High' reasoning settings via system prompts, allowing users to balance speed and detail.
- Tool Use: Excellent for web browsing with built-in tools, function calling with defined schemas, and agentic operations.
- Fine-tuning: The 20B parameter model can be fine-tuned on consumer hardware.
Use Cases
- Coding: Ideal for code generation, problem-solving, and synthesis.
- Complex Reasoning: Benefits from its 'High' reasoning setting for deep analysis tasks.
- Agentic Applications: Suitable for tasks requiring tool use, such as web browsing and function calling.
Guardrails
For enhanced safety, users are advised to set reasoning = "high" to prevent jailbreaking and prompt injection. Additionally, integrating openai/gpt-oss-safeguard-20b is recommended for robust guardrails.