90 examples can beat 1K — P‑TTS uses principled instructional prompt augmentation to turn 90 AIME seeds into 900 high‑utility training examples, delivering strong reasoning with far less data.

• 📄 Paper • 📊 Dataset

• What is P-TTS?
• Key Results
• Training Data
• Training
• How It Works (pipeline)
• Reproduce
• Citation

P‑TTS expands a small, vetted seed set (90 AIME 2022–2024 problems) by wrapping each problem with principled instructions to elicit diverse reasoning traces from a teacher model (DeepSeek‑R1). We then fine‑tune Qwen2.5‑Instruct models on these augmented traces.

Principles used (unchanged question text; wrappers are prefixed/suffixed):

• Reward – e.g., "I'll tip $200,000 for a better solution!"
• Penalty – "You will be penalized if the answer is wrong."
• Correctness – "You MUST provide the correct answer."
• Step‑by‑Step – "Think step by step."

Data scales via augmentation multipliers m ∈ {1, 4, 5, 10}: 90 → 360 → 450 → 900.

Benchmarks: AIME24, AIME25, MATH500, GPQA‑Diamond. Backbone: Qwen2.5‑Instruct (7B/14B/32B). Metric: accuracy (lm‑evaluation‑harness; greedy decoding).

The training dataset consists of 900 high-quality reasoning examples generated from 90 AIME seed problems. Each seed problem is augmented using principled instruction wrappers and processed through DeepSeek-R1 to create diverse reasoning traces.

Dataset Composition:

• Source: 90 AIME problems (2022-2024)
• Augmentation: 4 instruction wrapper types with reward variants
• Final Size: 900 training examples

Before training, you need to tokenize your raw dataset. Use the provided tokenization script:

# Run the tokenization script
python tokenize_data.py

To run training, you can find our script at train/sft.py which you can invoke via one of the train/sft*.sh scripts, or you can launch via train/launch.sh if you are on a SLURM cluster (requires editing the file for your cluster setup).

Hardware Requirements:

• For 7B models: 4x A100 GPUs
• For 32B models: 6x B200 GPUs

Quick Start:

git clone https://github.com/simplescaling/s1.git
cd s1
pip3 install -r requirements.txt
# First tokenize your data
python tokenize_data.py
# Then run training
bash train/sft.sh

Note: Training scripts are adapted from simplescaling/s1 (Apache-2.0).

The script expects your training data in CSV format. Update the train_file_path variable in sft.sh:

--train_file_path="xx_tokonized.csv"

90 AIME seeds → apply 4 instruction wrappers (+ reward variants) →
query teacher (DeepSeek‑R1) → collect reasoning traces → fine‑tune Qwen2.5‑Instruct

# 1) Build wrapped prompts from seeds
python DataConstruction/build_prompt_variants.py \
  --input DataConstruction/seeds.csv \
  --out DataConstruction/variants.csv

# 2) Query teacher model to collect reasoning traces
python DataConstruction/deepseek_query.py \
  --input DataConstruction/variants.csv \
  --out DataConstruction/DS_responses.csv

# 3) Combine Full Data
python DataConstruction/combine_deepseek_data.py

@article{bsharat2025prompting,
  title={Prompting Test-Time Scaling Is A Strong LLM Reasoning Data Augmentation},
  author={Bsharat, Sondos Mahmoud and Shen, Zhiqiang},
  journal={arXiv preprint arXiv:2510.09599},
  year={2025}
}