claude_mythos_distilled_25k

Claude Mythos Distilled 25K A high-quality synthetic supervised fine-tuning (SFT) dataset designed to train and fine-tune any LLM to mirror the capabilities, reasoning style, agentic behavior, and technical depth of Anthropic's Claude Mythos (distilled frontier model).** Dataset Summary **Size**: 25,000 high-quality examples **Format**: JSONL with chat `messages` (user/assistant pairs) + rich metadata **Categories** (balanced for general + specialized capability): Cybersecurity & Red Teaming: 7,000 examples (zero-days, exploits, defenses, EDR evasion, etc.) Advanced Coding & SWE: 5,500 examples (high-performance systems, memory safety, formal verification, production optimization) Mathematical & Logical Reasoning: 3,000 examples (proofs, complexity, PAC-Bayes, number theory) Agentic Planning & Long-Horizon Tasks: 3,500 examples (autonomous research agents, multi-agent swarms, self-improving systems) Scientific Analysis & Research: 2,500 examples (AlphaFold-style, quantum, CRISPR, astrophysics) General Expert QA (Mythos-style): 3,500 examples (alignment, governance, philosophy of frontier AI) **Total tokens (approx)**: ~12–15 million (dense, high-signal content) **Created**: May 2026 (synthetic, reproducible with seed 42) **Source**: Fully synthetic — generated to emulate the output distribution and cognitive style of Claude Mythos Preview (gated Anthropic frontier model known for autonomous exploit chaining, SWE-Bench leadership, and long-horizon agentic performance). **Not real Claude/Mythos outputs** (Mythos remains gated). Intended Use This dataset enables **distillation-style fine-tuning** of open-weight or smaller models (Llama-3/4, Qwen2.5, Mistral, Gemma, DeepSeek, etc.) to achieve significantly closer performance to frontier models in: Complex multi-step cybersecurity analysis and defensive engineering Production-grade, security-hardened, high-performance code generation Rigorous mathematical and scientific reasoning Autonomous agent design, planning, and se