gss1147's picture
Upload 2 files
bf04ded verified

Qwen 3.7 Max Thinking — Distilled Reasoning Dataset

5,000 high-quality, no-duplicate chain-of-thought reasoning traces for knowledge distillation, fine-tuning, or research. Each example contains a problem, a detailed step-by-step thinking trace (mirroring the Qwen 3.7 Max Thinking reasoning style), and a final answer.

Dataset Format

File: qwen3.7_max_thinking_dataset.jsonl
Format: JSON Lines (one JSON object per line)
Encoding: UTF-8 (ASCII-safe content — no special Unicode characters)

Schema

{
  "problem":        "The input question or task prompt",
  "thinking_trace": "Multi-step chain-of-thought reasoning trace with self-verification",
  "answer":         "The final concise answer"
}

Example

{
  "problem": "A car starts at 9 m/s and accelerates at 10 m/s^2 for 7s. Find distance and final velocity.",
  "thinking_trace": "Using kinematics equations:\nv_f = v_0 + a*t = 9 + 10*7 = 79 m/s\nd = v_0*t + 0.5*a*t^2 = 9*7 + 0.5*10*7^2\n  = 63 + 0.5*10*49 = 308.0 m\nCheck: v_avg = (9+79)/2 = 44.0 m/s, d = v_avg*t = 308.0 m [OK]",
  "answer": "Distance = 308.0m, Final velocity = 79 m/s"
}

Category Distribution

Category Count Percentage
Coding & Algorithms ~1,200 ~24%
Mathematics (algebra, quadratics, probability, geometry, sequences, combinatorics, stats) ~1,200 ~24%
Scientific Reasoning (physics, chemistry, biology) ~830 ~17%
Number Theory (primality, GCD, modular exponentiation) ~590 ~12%
Data Analysis (linear regression, trend estimation) ~590 ~12%
Agentic & Multi-Step Planning ~290 ~6%
Logic & Puzzles (liar puzzles, water jug, Monty Hall, Latin squares) ~270 ~5%
Commonsense Causal Reasoning ~10 <1%
Creative / Fermi Estimation ~10 <1%
Prompt Engineering / Agent Workflows ~10 <1%

Reasoning Style

The thinking traces follow the Qwen 3.7 Max Thinking paradigm:

  1. Problem decomposition — break the problem into manageable steps
  2. Formula recall — state the relevant equation or approach
  3. Step-by-step computation — perform each operation with intermediate values
  4. Self-verification — check the result via alternative method or substitution
  5. Final answer extraction — present the clean final answer

This style is optimized for knowledge distillation where a smaller student model learns to mimic the extended reasoning process of a larger teacher model.

Key Properties

Property Value
Total examples 5,000
Duplicate problems 0 (SHA-256 problem-level dedup)
Duplicate entries 0 (SHA-256 full-entry dedup)
Min thinking trace length 44 chars
Max thinking trace length 456 chars
Avg thinking trace length ~199 chars
File size ~1.76 MB
Output encoding UTF-8, ASCII-safe

Use Cases

  • Knowledge distillation — train a smaller model to produce step-by-step reasoning
  • Supervised fine-tuning (SFT) — teach chain-of-thought reasoning
  • Reasoning benchmark — evaluate model reasoning quality
  • Prompt engineering research — study structured reasoning patterns
  • Agent training data — multi-step planning and tool-use reasoning

Generation

The dataset was generated programmatically by generate_dataset.py using templated generators with randomized parameters across 10 reasoning domains. Each call produces a unique combination of numeric values, problem text, and computed outputs, then filters through a SHA-256 deduplication pipeline at both the entry and problem levels.

Regenerate or Extend

python generate_dataset.py

Modify the generator functions or weights in the script to adjust:

  • Category balance
  • Difficulty ranges
  • Problem types
  • Total count (default 5,000)