metadata
license: apache-2.0
Codette GPT-OSS-20B Training Dataset
Overview
This repository contains the structured training dataset used to fine-tune openai/gpt-oss-20b into a behaviorally conditioned architecture referred to as Codette.
The goal of this dataset is not personality injection or artificial sentience simulation.
The objective is structured behavioral conditioning across:
- Recursive reasoning (RC+ξ framework)
- Multi-perspective synthesis
- Governance-aware responses
- Natural response enhancement
- Cross-module architectural coherence
- Dynamic explanation depth scaling
This dataset is designed for LoRA-based fine-tuning of GPT-OSS-20B using 4-bit quantization (QLoRA).
Dataset File
codette_gptoss20b_master_v3.jsonl
Total Samples: ~5,000
Format: JSON Lines
Structure per entry:
{
"instruction": "...",
"input": "",
"output": "...",
"metadata": {
"category": "...",
"depth": "simple | intermediate | technical",
"module": "..."
}
}
Key Training Principles
1. Dynamic Explanation Scaling
The model is trained to automatically adjust explanation depth based on user query context:
Simple explanations for general audiences
Intermediate explanations for practitioners
Technical explanations for formal requests
2. Governance Stability
Examples reinforce:
Ethical constraint adherence
Refusal handling with clarity
No bypass of safety mechanisms
3. RC+ξ Recursive Reasoning
The dataset conditions structured reasoning concepts including:
Epistemic tension (ξ)
Recursive state evolution
Convergence behavior
Attractor dynamics
These are applied contextually rather than injected indiscriminately.
4. Natural Response Enhancement
Examples train the model to:
Avoid robotic phrasing
Avoid system markers or bracket artifacts
Maintain clarity without over-verbosity
5. Cross-Module Integration
Training includes architectural reasoning across components such as:
Recursive reasoning
Natural enhancement layer
Governance system
Adaptive learning behaviors
Intended Use
This dataset is intended for:
LoRA fine-tuning of GPT-OSS-20B
Architectural behavioral conditioning
Research into structured recursive reasoning systems
Controlled deployment experiments
Not Intended For
Claims of machine consciousness
Identity simulation
Misrepresentation of system capabilities
Replacement for safety-aligned governance models
Recommended Training Configuration
4-bit NF4 quantization
LoRA rank 32
3 epochs
Learning rate: 1e-4
Cosine scheduler
A100 GPU recommended
Author
Jonathan Harrison
Raiff1982
License
Specify license here (e.g., Apache 2.0, MIT, or research-only).