instruct5 / README.md
tamewild's picture
Update README.md
eeaf9df verified
|
Raw History Blame Contribute Delete
1.73 kB
---
dataset_info:
features:
- name: conversation
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 12674536.395147314
num_examples: 500
- name: test
num_bytes: 1951878.6048526864
num_examples: 77
download_size: 5332626
dataset_size: 14626415
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
license: apache-2.0
---
# Instruct5 (500 Zebra Puzzles)
## Overview
This synthetic micro-dataset consists of **577 pure logic deduction traces (5x5 zebra puzzles)**, split into **500 training examples** and **77 validation examples**. It features strictly logic grid puzzles and includes **zero mathematical data**.
It was created to demonstrate how fine-tuning small base models on dense, long-form logic traces can elicit substantial out-of-domain reasoning generalization (e.g., boosting performance on MATH-500 and AIME).
For full training details and reproduction scripts, see the [PCSS GitHub Repository](https://github.com/tamewild/pcss).
## Data Generation & Cost
The dataset was generated using `Qwen3-235B-A22B-Instruct-2507` (served in FP8 via DeepInfra in September 2025). We employed rejection sampling using a correctness verifier.
Total API usage for the entire generation process:
- **Total Cost:** $4.31 ($0.06 Input, $4.25 Output)
- **Input Tokens:** 434,379 ($0.13 / 1M tokens)
- **Output Tokens:** 7,075,393 ($0.60 / 1M tokens)
*(Note: These figures represent the total bill for this model that month. They serve as a strict upper bound, capturing all test runs and discarded samples rather than just the tokens in the final dataset.)*