barozp/opus-reasoning-distill-train

Claude Opus Reasoning Distillation Dataset

Claude Opus 4.6/4.7 reasoning traces, formatted for SFT fine-tuning. This repo is the train split.

Dataset size

Purpose

Distill Claude Opus's structured reasoning style into Qwen3.6-27B (Bonsai base model).

Data Sources

Sources were pooled, standardized to a single messages format, shuffled, and split 95/5 into 14,250 train + 750 validation (15,000 total).

Format

Each example uses the messages format with <think>...</think> reasoning blocks:

{
  "messages": [
    {"role": "user", "content": "How many positive integers less than 1000 have digits that sum to 20?"},
    {"role": "assistant", "content": "<think>\nI need to count positive integers...\n</think>\n\n**36** positive integers."}
  ]
}

Usage

Designed for SFT with train_on_responses_only to train only the assistant/thinking portion:

from datasets import load_dataset

dataset = load_dataset("barozp/opus-reasoning-distill-train", split="train")
# Apply Qwen chat template, train with Unsloth + SFTTrainer

Preparation Script

prepare_dataset.py — Loads all sources, standardizes to messages format, shuffles, and splits 95/5.