Caveman Reasoning Compression
Collection
Compressed (caveman) reasoning traces train models to reason in ~2-3x fewer tokens at equal or better accuracy. For full ablation: sibling collection. • 9 items • Updated
prompt_id stringlengths 16 16 | legacy_id stringlengths 34 52 ⌀ | prompt stringlengths 32 12.5k | source stringclasses 2
values | lang stringclasses 1
value | task_type stringclasses 26
values | domain stringclasses 4
values | reasoning stringlengths 396 42.3k | response stringlengths 11 40k | verified bool 1
class | verify_method stringclasses 3
values | difficulty dict | degenerate bool 2
classes | gen dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
8d481218e8a4e165 | PrimeIntellect/SYNTHETIC-2-SFT-verified_3ebd563ac37c | Wrap your entire response with double quotation marks. Generate a list of three social issues Output: | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | ifeval | general | Thinking Process:
1. **Analyze the Request:**
* Task: Generate a list of three social issues.
* Constraint: Wrap the *entire* response with double quotation marks.
* Output Format: "List of three social issues..."
2. **Identify the Content:**
* I need to select three common social issues. Ex... |
"1. Income Inequality
2. Climate Change
3. Access to Healthcare" | null | none | null | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 1
} |
211b014398b377b9 | PrimeIntellect/SYNTHETIC-2-SFT-verified_8deae09d4af6 | You are given an n x n grid where each cell can have one of three values:
- 0 representing an empty cell
- 1 representing a fresh orange
- 2 representing a rotten orange
Every minute, any fresh orange that is 4-directionally adjacent to a rotten orange becomes rotten.
Your task is determine the minimum number of minu... | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | reasoning_gym | general | The problem asks for the minimum number of minutes until no fresh oranges (value 1) remain in an $n \times n$ grid.
The grid is given as:
$$
\begin{bmatrix}
1 & 1 & 1 \\
0 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix}
$$
Values:
- 0: Empty
- 1: Fresh
- 2: Rotten
This is a classic Breadth-First Search (BFS) problem (LeetCode 994:... |
```json
{
"answer": -1
}
``` | null | none | null | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 1
} |
9815a3616820390e | PrimeIntellect/SYNTHETIC-2-SFT-verified_8b24dec2e6ae | Give two different responses. Responses and only responses should be separated by 6 asterisk symbols: ******. Question:
Основные возможности: "Чистое" удаление программ, быстрее в 3 раза чем стандартная Установка и Удаление программ, удаление программ, автоматически запускаемых при старте системы
Could you please tran... | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | ifeval | general | Thinking Process:
1. **Analyze the Request:**
* Input: A series of examples (Question + Answer) showing Russian text being translated to English.
* Task: Provide a translation for the last question: "Вы так же можете скачать перевод текста песни Thunderstone Eyes Of A Stranger здесь . Мы стараемся сделать... |
You can also download the translation of the song lyrics for Thunderstone's "Eyes Of A Stranger" here. We strive to ensure that the words of the song "Eyes Of A Stranger" are as accurate as possible, so if you have any corrections to the text, please send them to us.
******
You may also download the lyrics translatio... | null | none | null | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 1
} |
193167b5ef01f492 | PrimeIntellect/SYNTHETIC-2-SFT-verified_e86af45c628e | AtCoder Inc. holds a contest every Saturday.
There are two types of contests called ABC and ARC, and just one of them is held at a time.
The company holds these two types of contests alternately: an ARC follows an ABC and vice versa.
Given a string S representing the type of the contest held last week, print the str... | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | prime_rl_code | code | Thinking Process:
1. **Analyze the Request:**
* Problem: Given a string $S$ representing the contest type held last week (either "ABC" or "ARC"), determine the contest type held this week.
* Rule: Contests alternate. If last week was ABC, this week is ARC. If last week was ARC, this week is ABC.
* I... |
To solve this problem, we need to determine the contest type for the current week based on the contest type held last week. The problem states that the contests alternate between "ABC" and "ARC".
Here is the step-by-step reasoning:
1. **Read Input**: We read the string $S$ from Standard Input, which represents the ... | true | exec | {
"solve_rate": 1,
"k": 4
} | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
909b3b2fb8dd9200 | PrimeIntellect/SYNTHETIC-2-SFT-verified_77ce694b3821 | 6. For all non-negative values of the real variable $x$, the function $f(x)$ satisfies the condition $f(x+1)+1=f(x)+\frac{43}{(x+1)(x+2)}$. Calculate $\frac{101}{f(2020)}$, if $\quad f(0)=2020$. | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | verifiable_math | math | Here's a thinking process to solve the problem:
1. **Understand the Goal:** The objective is to calculate the value of $\frac{101}{f(2020)}$, given a functional equation for $f(x)$ and an initial condition $f(0) = 2020$.
2. **Analyze the Given Information:**
* Domain: $x \ge 0$.
* Functional Equation: $... |
The given functional equation is:
$$f(x+1) + 1 = f(x) + \frac{43}{(x+1)(x+2)}$$
Rearranging the terms to express $f(x+1)$ in terms of $f(x)$:
$$f(x+1) = f(x) + \frac{43}{(x+1)(x+2)} - 1$$
We are interested in $f(2020)$, which corresponds to integer values. Let $n$ be a non-negative integer. We can write $f(n)$ as a s... | true | symbolic | {
"solve_rate": 0.25,
"k": 4
} | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
9a1ea1061b87b416 | PrimeIntellect/SYNTHETIC-2-SFT-verified_c23602b9fc47 | (7) Given the sequence $\left\{a_{n}\right\}$ where the first term $a_{1}=2, a_{n+1}=\frac{5 a_{n}-13}{3 a_{n}-7}$, then $a_{2011}=$ | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | verifiable_math | math | The user wants to find the value of $a_{2011}$ given the sequence defined by $a_1 = 2$ and $a_{n+1} = \frac{5 a_n - 13}{3 a_n - 7}$.
Let's compute the first few terms to see if there's a pattern.
$a_1 = 2$.
$a_2 = \frac{5(2) - 13}{3(2) - 7} = \frac{10 - 13}{6 - 7} = \frac{-3}{-1} = 3$.
$a_3 = \frac{5(3) - 13}{3(3) - 7... |
The sequence $\{a_n\}$ is defined by the recurrence relation $a_{n+1} = \frac{5 a_n - 13}{3 a_n - 7}$ with the initial term $a_1 = 2$.
Let's compute the first few terms of the sequence to identify a pattern.
For $n=1$:
$a_1 = 2$.
$a_2 = \frac{5(2) - 13}{3(2) - 7} = \frac{10 - 13}{6 - 7} = \frac{-3}{-1} = 3$.
For $n... | true | symbolic | {
"solve_rate": 0.75,
"k": 4
} | true | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
209c563ad06b3954 | PrimeIntellect/SYNTHETIC-2-SFT-verified_903ee4b7b71a | Please critique whether the following solution to the question is correct.
Question:
Prove that the complex function \( f(z) = \frac{1+z}{1-z} \) maps the interior of the unit circle \( |z| < 1 \) in the \( z \) plane to the right half-plane \( \Re f(z) > 0 \) in the \( f(z) \) plane.
Solution:
To demonstrate this, c... | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | no_verification | general | Here's a thinking process that leads to the critique:
1. **Analyze the Request:**
* **Task:** Critique whether the provided solution to a specific complex analysis problem is correct.
* **Question:** Prove that $f(z) = \frac{1+z}{1-z}$ maps the interior of the unit circle ($|z| < 1$) to the right half-pla... |
The solution provided is **incorrect**. While it correctly identifies the inverse mapping function and performs a valid inequality check for the magnitude, it contains a fundamental algebraic and logical error in the second half, and it lacks necessary rigor in the first half.
Here is a detailed critique broken down... | null | none | null | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 1
} |
261f0ef31d68b918 | PrimeIntellect/SYNTHETIC-2-SFT-verified_2dd5d138d7da | G3.3 Given that for any real numbers $x, y$ and $z, \oplus$ is an operation satisfying
(i) $x \oplus 0=1$, and
(ii) $\quad(x \oplus y) \oplus z=(z \oplus x y)+z$.
If $1 \oplus 2004=c$, find the value of $c$. | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | verifiable_math | math | The problem asks us to find the value of $c = 1 \oplus 2004$, given the properties of an operation $\oplus$ on real numbers.
The properties are:
(i) $x \oplus 0 = 1$ for all real numbers $x$.
(ii) $(x \oplus y) \oplus z = (z \oplus x y) + z$.
Let's denote the operation $\oplus$ as $x \oplus y$.
We are given $c = 1 \op... |
The problem asks for the value of $c = 1 \oplus 2004$, given the properties of the operation $\oplus$ defined on real numbers:
1. $x \oplus 0 = 1$ for any real number $x$.
2. $(x \oplus y) \oplus z = (z \oplus xy) + z$ for any real numbers $x, y, z$.
We want to find $1 \oplus 2004$. Let's analyze the second condit... | true | symbolic | {
"solve_rate": 0.25,
"k": 4
} | false | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
b95ec8d361e30744 | PrimeIntellect/SYNTHETIC-2-SFT-verified_54791a588cd4 | What is $\frac{2+4+6}{1+3+5} - \frac{1+3+5}{2+4+6}?$ | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | verifiable_math | math | Here's my thought process for solving this expression:
1. **Analyze the Request:** The user wants me to evaluate a mathematical expression: $\frac{2+4+6}{1+3+5} - \frac{1+3+5}{2+4+6}$. The final answer should be boxed.
2. **Break Down the Expression:**
* The expression consists of two fractions being subtract... |
To find the value of the expression $\frac{2+4+6}{1+3+5} - \frac{1+3+5}{2+4+6}$, we first evaluate the sums in the numerators and denominators.
1. **Calculate the sums:**
* The sum in the first numerator is $2 + 4 + 6 = 12$.
* The sum in the first denominator (and second numerator) is $1 + 3 + 5 = 9$.
... | true | symbolic | {
"solve_rate": 0.75,
"k": 4
} | true | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
bbee5c279b267a9b | PrimeIntellect/SYNTHETIC-2-SFT-verified_05a5d9f12e15 | Taro had his own personal computer and set a password for login. However, Taro inadvertently forgot the password. Then, remembering that there was a piece of paper with the password written down, Taro found the paper and was surprised to see it. The paper was cut and there were only fragments, and there were some stain... | PrimeIntellect/SYNTHETIC-2-SFT-verified | en | prime_rl_code | code | The user wants a Python program to solve a specific problem.
**Problem Description:**
Taro has a password string $A$ and a memo string $B$.
$B$ is a fragment of $A$ with some characters potentially unreadable (represented by underscores `_`).
The goal is to determine if $B$ can be a valid fragment of $A$.
Actually, lo... |
The problem asks us to determine if string $B$ is a "fragment" of string $A$. Based on the examples and the description of the paper being cut and having stains (represented by `_`), the problem reduces to checking if string $B$ is a substring of string $A$ where the characters in $B$ that are not `_` must match the ... | true | exec | {
"solve_rate": 0.5,
"k": 4
} | true | {
"model": "Qwen/Qwen3.5-4B",
"temperature": 0.6,
"top_p": 0.95,
"max_tokens": 12288,
"k": 4
} |
End of preview. Expand in Data Studio
crucible-sft-qwen3.5-4b-mini
Self-distilled SFT dataset of verified reasoning traces from Qwen/Qwen3.5-4B, built by the reasoning-compression crucible pipeline: k-sample generation on a decontaminated prompt pool, inline verification (symbolic math / sandboxed code tests), difficulty banding via solve rate, and loop-detector filtering on the chosen trace.
Each row: prompt, reasoning (a verified-correct thinking trace when the
domain is verifiable), response, domain, verified, difficulty
({solve_rate, k}), sampling params in gen.
- Rows: 648 — domains: {"general": 219, "code": 101, "math": 191, "creative": 137}
- Prompts are content-addressed (
prompt_id= sha256 of prompt text) and decontaminated against GSM8K, MATH-500, HumanEval, MBPP, MMLU-Pro. - A stratified held-out split was carved before generation and is not included here.
Part of the caveman reasoning-compression ablation study (see the flint datasets built from this file and the trained comparison arms).
Build manifest
{
"created_at": "2026-07-09T11:20:19+00:00",
"git_sha": "4636916a99c191bd15795c43fbee26ac3337d1f5",
"recipe": "flint-ablation-v2-mini",
"recipe_def": {
"target_size": 1100,
"shares": {
"math": 0.3,
"code": 0.16,
"general": 0.28,
"creative": 0.26
},
"candidate_caps": {
"math": 700,
"code": 354,
"general": 400,
"creative": 350
},
"solve_band": [
0.2,
0.8
],
"solve_band_by_domain": {
"code": [
0.2,
1.0
]
},
"k": 4,
"oversample": 2.2,
"lang": "en"
},
"raw_rows": 1328,
"selected_rows": 648,
"domains": {
"general": 219,
"code": 101,
"math": 191,
"creative": 137
},
"verified": {
"None": 356,
"True": 292
},
"degenerate_in_raw": 272,
"gen_model": "Qwen/Qwen3.5-4B"
}
- Downloads last month
- 23