Datasets:

wrmedford
/

Gemma-4-E4B-it-SSD

messages listlengths 2 2
[ { "role": "user", "content": "Sorting arrays is traditionally associated with high-level languages. How hard can it be in Roco? Sort the given array in non-descending order.\n\nInput\n\nThe first line of input contains an integer n (1 ≤ n ≤ 100) — the size of the array. The following n lines contain the ele...
[ { "role": "user", "content": "You are given an equation: \n\nAx2 + Bx + C = 0. \n\nYour task is to find the number of distinct roots of the equation and print all of them in ascending order.\n\nInput\n\nThe first line contains three integer numbers A, B and C ( - 105 ≤ A, B, C ≤ 105). Any coefficient may be...
[ { "role": "user", "content": "In computer science, there is a method called \"Divide And Conquer By Node\" to solve some hard problems about paths on a tree. Let's desribe how this method works by function:\n\nsolve(t) (t is a tree): \n\n 1. Chose a node x (it's common to chose weight-center) in tree t. Le...
[ { "role": "user", "content": "A recently found Ancient Prophesy is believed to contain the exact Apocalypse date. The prophesy is a string that only consists of digits and characters \"-\".\n\nWe'll say that some date is mentioned in the Prophesy if there is a substring in the Prophesy that is the date's re...
[ { "role": "user", "content": "There are n cows playing poker at a table. For the current betting phase, each player's status is either \"ALLIN\", \"IN\", or \"FOLDED\", and does not change throughout the phase. To increase the suspense, a player whose current status is not \"FOLDED\" may show his/her hand t...
[ { "role": "user", "content": "Advertising has become part of our routine. And now, in the era of progressive technologies, we need your ideas to make advertising better!\n\nIn this problem we'll look at a simplified version of context advertising. You've got a text, consisting of exactly n words. A standard...
[ { "role": "user", "content": "Everybody knows that we have been living in the Matrix for a long time. And in the new seventh Matrix the world is ruled by beavers.\n\nSo let's take beaver Neo. Neo has so-called \"deja vu\" outbursts when he gets visions of events in some places he's been at or is going to be...
[ { "role": "user", "content": "We know that lucky digits are digits 4 and 7, however Vasya's got another favorite digit 0 and he assumes it also is lucky! Lucky numbers are such non-negative integers whose decimal record only contains lucky digits. For example, numbers 0, 47, 7074 are lucky, but 1, 7377, 895...
[ { "role": "user", "content": "Vasily the Programmer loves romance, so this year he decided to illuminate his room with candles.\n\nVasily has a candles.When Vasily lights up a new candle, it first burns for an hour and then it goes out. Vasily is smart, so he can make b went out candles into a new candle. A...
[ { "role": "user", "content": "Petya studies positional notations. He has already learned to add and subtract numbers in the systems of notations with different radices and has moved on to a more complicated action — multiplication. To multiply large numbers one has to learn the multiplication table. Unfortu...
[ { "role": "user", "content": "Sereja has painted n distinct points on the plane. The coordinates of each point are integers. Now he is wondering: how many squares are there with sides parallel to the coordinate axes and with points painted in all its four vertexes? Help him, calculate this number.\n\nInput\...

End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SSD Dataset Replication (Gemma-4-E4B-it)

This dataset is a replication of the "Embarrassingly Simple Self-Distillation Improves Code Generation" (SSD) paper (arXiv:2604.01193).

Overview

The dataset contains coding problems and their corresponding solutions generated by Gemma-4-E4B-it using high-temperature sampling (T=1.1) to explore the model's latent capabilities. This approach, known as SSD, focuses on "self-distillation" where a model's own correct but non-greedy outputs are used for fine-tuning to improve its standard (greedy) performance.

Dataset Construction

Seed Prompts: Samples from deepmind/code_contests.
Generation: High-temperature sampling (T=1.1, Top-K=20, Top-P=0.95) with Gemma-4-E4B-it.
Reasoning: Reasoning chains (thinking) were enabled during generation to capture the model's logical process.

ssd_dataset.jsonl: The generated samples (13,328 entries) in OpenAI-style message format.
generate_dataset.py: The replication script used to synthesize the data (vLLM-based, data-parallel).

Usage

To replicate the generation process:

python generate_dataset.py --model google/gemma-4-E4B-it --dp <NUM_GPUS>

Citation

@misc{zhang2026embarrassinglysimpleselfdistillationimproves,
      title={Embarrassingly Simple Self-Distillation Improves Code Generation}, 
      author={Ruixiang Zhang and Richard He Bai and Huangjie Zheng and Navdeep Jaitly and Ronan Collobert and Yizhe Zhang},
      year={2026},
      eprint={2604.01193},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.01193}, 
}