kaushik-harsh-99/math-sft-solutions-no-cot-v3

Math SFT Solutions No CoT V3

Math SFT Solutions No CoT V3 is a large-scale mathematics supervised fine-tuning (SFT) dataset designed for instruction tuning and mathematical capability adaptation.

Version 3 substantially expands mathematical coverage while improving dataset quality through stronger filtering, cleaning, and supervision refinement.

Unlike reasoning-heavy datasets, this release focuses on clean instruction → response pairs without hidden chain-of-thought style supervision.

Dataset Summary

This dataset provides concise mathematical supervision for training and adapting language models.

Dataset format:

{
  "instruction": "...",
  "response": "..."
}

Features

• —Broad mathematical domain coverage
• —Instruction → response format
• —No hidden reasoning traces
• —Deduplicated and cleaned
• —Synthetic augmentation included
• —Optimized for supervised fine-tuning
• —Suitable for mathematical adaptation

What's New in Version 3

Expanded Mathematical Coverage

Version 3 significantly increases mathematical diversity and domain coverage.

Included domains:

• —Arithmetic
• —Pre-Algebra
• —Algebra
• —Geometry
• —Trigonometry
• —Calculus
• —Number Theory
• —Probability
• —Statistics
• —Combinatorics
• —Discrete Mathematics
• —Symbolic Manipulation
• —Competition Mathematics
• —Multi-step Problem Solving
• —Mixed Difficulty Mathematical Reasoning

The objective is broader mathematical supervision and improved generalization.

Cleaner Supervision Targets

Version 3 continues the cleanup introduced in Version 2.

Processing improvements include:

• —removal of reasoning artifacts
• —removal of thinking blocks
• —response normalization
• —formatting cleanup
• —duplicate removal
• —near-duplicate filtering
• —quality-oriented preprocessing

Responses contain only intended supervised targets.

This dataset is not intended for hidden chain-of-thought supervision.

Increased Diversity

Version 3 expands beyond earlier GSM8K-style and MATH-style distributions.

Data construction emphasizes:

• —wider mathematical structures
• —varied instruction styles
• —broader solution distributions
• —multiple difficulty ranges

Benchmark improvements are not guaranteed and depend on training setup.

Dataset Structure

Data Fields

Example

{
  "instruction": "Solve for x: 3x + 9 = 24",
  "response": "Subtract 9 and divide by 3. Final answer: x = 5."
}

Data Sources

This dataset contains transformed and augmented mathematical examples derived from:

• —GSM8K-style arithmetic tasks
• —MATH-style mathematics tasks
• —synthetic mathematical augmentation pipelines
• —transformed instruction–response datasets
• —expanded multi-domain mathematical supervision

Coverage includes:

• —Arithmetic
• —Algebra
• —Geometry
• —Calculus
• —Number Theory
• —Combinatorics
• —Symbolic Manipulation
• —Mathematical Reasoning

Intended Uses

Recommended for:

• —Supervised Fine-Tuning (SFT)
• —Instruction Tuning
• —Mathematical Adaptation
• —LoRA
• —QLoRA
• —Response Generation
• —Small Model Specialization

Compatible with:

• —Qwen
• —Llama
• —Gemma
• —SmolLM
• —Mistral
• —other decoder-only language models

Limitations

• —contains synthetic augmentation
• —mathematical correctness is not guaranteed for every sample
• —not intended for theorem verification
• —not intended for hidden chain-of-thought training
• —benchmark performance depends on training setup
• —may contain residual distribution artifacts

Version History

V1

Initial release.

Contained intermediate reasoning-format contamination.

V2

Introduced:

• —removal of thinking contamination
• —cleaner supervision targets
• —response augmentation

V3

Introduced:

• —expanded mathematical coverage
• —stronger preprocessing
• —broader supervision
• —improved filtering
• —cleaner instruction tuning targets