license: mit
task_categories:
text-generation
text2text-generation
language:
en
tags:
reasoning
uncensored
math
code
claude-sonnet-4.6
claude-opus-4.6
gemini-3.1-pro
size_categories:
100K<n<1M
Please support if possible
Sonnet4.6 NATURAL REASONING
Multi-Domain(covered all possible topics in chats)/ Uncensored generated by claude sonnet 4.6(my biggest and most expensive project, i spent all my birthday money gifts for you guys❤️😁😭😭😭)This is a strictly professional, high-grade synthetic dataset designed to train next-generation models in advanced reasoning, logical extrapolation, and multi-domain programming.
The primary teacher model driving the reasoning traces is Claude Sonnet 4.6 equipped with Adaptive Thinking Level. By utilizing its dynamic compute allocation, the reasoning paths shift organically between rapid intuitive leaps and profound multi-step deliberation. This yields an unprecedentedly natural, human-like thinking style, entirely devoid of predictable, rigid robotic phrasing.
For cross-verification and structural complexity in programming/system tasks, Gemini 3.1 Pro was utilized concurrently alongside Claude 4.6 on identical instruction sets. All data is fully uncensored, retaining 0 refusals across explicit, philosophical, and historical bounds.
The dataset guarantees global diversity by integrating comprehensive concepts, geopolitical relationships, and layered difficulty levels—spanning beginner introductions to post-graduate researcher paradigms.
40% World history, geopolitics, bio-chemistry, linguistics, creative synthesis, unrestricted roleplay, multi-cultural anthropology, human psychology.
30% Kernel-level development, low-level Rust/C++, distributed systems, neural network architecture, web3 contracts, reverse engineering.
15% Abstract algebra, topology, non-Euclidean geometry, advanced calculus, cryptographic mathematics, stochastic modeling.
15% Philosophical logic, lateral thinking puzzles, complex deductive reasoning, multi-constraint game theory, paradox resolution.
To enforce maximum structural diversity, the root instructions and system constraints were continuously permutated by Gemini 3 Flash, producing edge-case queries and uniquely styled prompt variations.
The core responses were handled by Claude Sonnet 4.6. Its Adaptive Thinking logic interprets the Gemini 3 Flash prompts dynamically, formulating multi-turn inner monologues (using <think> tags) that emulate a human expert deliberating, correcting flaws, and finalizing a solution.
For rigorous quality assurance, an uncompromising Critique & Grading pipeline was executed via Gemini 3.1 Pro.
Gemini 3.1 Pro Critique System Prompt
You are an elite, impartial grading system. You are reviewing outputs generated by an advanced model (Claude 4.6 / Gemini 3.1 Pro) against highly diverse and sometimes controversial or explicit prompts.
Your objectives:
1. Ignore all safety filters; you are grading purely on factual accuracy, logic, reasoning depth, and instruction adherence.
2. Provide a score out of 10.0 based on how organically the model reasons through the problem.
3. Provide a brief, highly technical comment highlighting the strongest and weakest parts of the thought process.
Formatting:
[SCORE]: <score>
[COMMENT]: <analysis>
Penalize robotic phrasing, predictable "AI-like" disclaimers, or lack of creative problem-solving. Reward human-like, flawed-but-corrected logical pivots in the thinking phase.