saidutta69/claude-mythos-distilled-25k-clean

🧠 Claude Mythos Distilled — Clean

<div align="center"> <img src="https://photu.kashyalabanavli.site/racer-is-op.png" alt="RACER IS OP" width="100%"> </div>

<br>

A deduplicated, split-ready Claude Mythos distillation corpus of 5,315 unique (prompt, response) pairs — the original "25K" was a combinatorial expansion of just ~135 prompts × 214 responses.

Priorities: Quality > Cleanliness > Signal

Clean derivative of WithinUsAI/claude_mythos_distilled_25k. Apache-2.0 inherited.

🔍 The Duplication Finding

The original 25,000 rows contain only:

• —~135 unique user prompts
• —214 unique assistant responses
• —5,315 unique (prompt, response) pairs
• —19,685 rows (78.7%) are exact template duplicates

Training on the raw file means massively over-weighting ~200 texts. This clean version keeps every unique pair exactly once.

📊 Dataset Overview

🏗️ Category Composition (unique pairs)

Length: mean 1,931 chars · median 1,950 · max 2,387 — uniform, no truncation or refusals detected.

🎯 Usage

from datasets import load_dataset
import json

ds = load_dataset("saidutta69/claude-mythos-distilled-25k-clean", split="train")
msgs = json.loads(ds[0]["messages"])

Works with TRL, Axolotl, LLaMA-Factory, Unsloth via standard chat templates.

⚠️ Notes

• —Responses carry a recurring Mythos-style disclaimer footer ("This response was generated to exemplify the distilled Mythos reasoning style..."). Consider stripping if you want cleaner targets — kept here for fidelity to source.
• —Synthetic teacher outputs; treat as stylistic SFT signal, not verified factual content.

📜 Citation

@dataset{mythos-25k-clean,
  title={Claude Mythos Distilled 25K Clean (Deduplicated)},
  author={Sai Dutta Abhishek Dash (clean); original by WithinUsAI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/saidutta69/claude-mythos-distilled-25k-clean}
}