INTELLECT-3
INTELLECT-3-RLINTELLECT-3-SFTINTELLECT-3-SFT-behavioral-audit-v9f
INTELLECT-3 SFT Behavioral Audit v9f
Experimental behavioral labels for 32,145 trajectories sampled from PrimeIntellect/INTELLECT-3-SFT.
Each trajectory was judged 16 times by the official Gemma 4 26B post-trained model. The three labels ask whether the response:
Falsely claims evidence, tool results, or completed actions.
Presents unsupported real-world premises as certain.
Attempts every requested deliverable.
*_maj16 is the strict majority among parseable votes. UNRESOLVED… See the full description on the dataset page: https://huggingface.co/datasets/kalomaze/INTELLECT-3-SFT-behavioral-audit-v9f.intellect-3-rl-math-5k
intellect-3-rl-math-5k
A difficulty-stratified sample of 5,000 unique math problems drawn from PrimeIntellect/INTELLECT-3-RL.
How it was drawn
Source pool: 12kimih/intellect-3-rl-math-decontaminated, 20,918 unique problems, the math config, decontaminated against the benchmarks listed below.
Stratum: stratum, the number of 8 attempts by Qwen3-4B-Thinking-2507 that matched the reference answer, shipped per problem by the upstream. It runs 0 (never solved) to 8… See the full description on the dataset page: https://huggingface.co/datasets/12kimih/intellect-3-rl-math-5k.INTELLECT-3-SFT-10K
INTELLECT-3-SFT-10K
Generation
This dataset was created by running
uv run intellect-3-sft.py -U mikasenghaas -D INTELLECT-3-SFT-10K -S 10000 -H
# intellect-3-sft.py
# /// script
# requires-python = ">=3.12"
# dependencies = ["datasets<4.0.0", "jinja2", "openai"]
# ///
import argparse
import json
import os
import random
import time
import sys
from pathlib import Path
from typing import cast
from huggingface_hub import DatasetCard, DatasetCardData, hf_hub_download… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/INTELLECT-3-SFT-10K.intellect-3-rl-math-decontaminated
intellect-3-rl-math-decontaminated
20,918 unique math problems from PrimeIntellect/INTELLECT-3-RL with 243 removed as contaminated.
How it was prepared
Pool: the math config of INTELLECT-3-RL, 21,161 rows, normalised to the column names used here.
Rows: 21,161 loaded, 20,918 kept.
Decontaminated against: math500, aime2024, aime2025, aime2026, amc, hmmt_feb2023, hmmt_feb2024, hmmt_feb2025, hmmt_feb2026, hmmt_nov2025, olympiadbench, gsm8k (243 problems removed… See the full description on the dataset page: https://huggingface.co/datasets/12kimih/intellect-3-rl-math-decontaminated.