yrdu/S3-CoT-Self-Sampled-Data

Dataset Overview

We release a self-sampled variable-length reasoning dataset collected from four models.
The data are used for efficient cot learning in our paper (S3-CoT: Self-Sampled Succinct Reasoning Enables Efficient Chain-of-Thought LLMs).
For each instruction (problem), we provide multiple reasoning traces generated under different activation-steering ratios, together with token-length statistics and answer extraction based on \boxed{} parsing.
This dataset is designed to support research on efficient reasoning, reasoning-length control, and length–accuracy trade-offs.

Source Models and Files

We have published self-sampled variable-length data based on four different models.

Data Schema

Each example is a JSON object with the following fields:

• instruction: The input prompt / question.
• gold_answer: The reference answer.

Multi-ratio reasoning traces

For each instruction, we provide reasoning traces (model outputs) sampled under multiple activation steering strengths:

• chain_0.0, chain_-0.1, chain_-0.3, chain_-0.5: The generated responses (typically containing a final answer in \boxed{}).
• ratios_0.0, ratios_-0.1, ratios_-0.3, ratios_-0.5: The ratio of the sampled length to the initial length (ratios_0.0 is fixed to 1.0).

Token-length statistics

• tokens_len_0.0, tokens_len_-0.1, tokens_len_-0.3, tokens_len_-0.5: Total token counts (all_tokens) of each trace.

Answer extraction and correctness

We extract the final numeric answer from each response by parsing the content inside \boxed{...}:

• answer_0.0, answer_-0.1, answer_-0.3, answer_-0.5: Extracted boxed answers.
• is_correct_0.0, is_correct_-0.1, is_correct_-0.3, is_correct_-0.5: Whether the extracted answer matches gold_answer.

Availability flag

To ensure compatibility with common context limits and to filter overly short/degenerate outputs, we include an availability indicator:

• is_available_0.0, is_available_-0.1, is_available_-0.3, is_available_-0.5: True if the trace length satisfies 100 < all_tokens < 4095, otherwise False.

Intended Use

This dataset can be used for:

• studying reasoning-length control and length–accuracy trade-offs,
• training or evaluating models that internalize concise reasoning,
• analyzing how steering ratios affect intermediate reasoning trajectories and final answers.