vanek-epfl/tulu3-sft-mixture-sampled

Tulu 3 SFT Mixture (Sampled)

This dataset is a sampled and filtered subset of the `allenai/tulu-3-sft-mixture`, curated and rebalanced for structured instruction fine-tuning. The goal is to support research and model development in math reasoning, coding, knowledge recall, instruction following (IF), and conversational alignment, while explicitly excluding safety, multilingual, and certain task-specific sources.

📦 Dataset Structure

• —Source: Filtered from allenai/tulu-3-sft-mixture
• —Size: 450,000 total samples
• —Splits:

• —train: 425,000 samples
• —test: 25,000 samples
• —Format: Hugging Face DatasetDict with "source", "prompt", "response" fields

🎯 Sampling Goals

The dataset is proportionally sampled to cover five main instruction types:

🧹 Excluded Categories

The following datasets were filtered out for alignment and safety reasons:

• —Safety: coconot, wildjailbreak, wildguardmixtrain
• —Multilingual: aya
• —Structured / Format-Specific Tasks: table_gpt, tulu_hard_coded_repeated
• —Style / Format Divergence: no_robots

🗂️ Dataset Categories & Sources

Below is the breakdown of included sources by category:

🧮 Math

• —personahub_math_v5_regen_149960
• —numinamath_tir_math_decontaminated
• —tulu_v3.9_open_math_2_gsm8k_50k
• —tulu_v3.9_personahub_math_interm_algebra_20k
• —tulu-3-sft-personas-math-grade

🧑‍💻 Coding

• —personahub_code_v2_34999
• —evol_codealpaca_heval_decontaminated

🧠 Knowledge Recall

• —flan_v2_converted
• —tulu_v3.9_sciriff_10k

💬 Chat

• —tulu_v3.9_wildchat_100k
• —oasst1_converted

🔁 Instruction Following (IF)

• —personahub_ifdata_manual_seed_v3_29980

🚀 Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("vanek-epfl/tulu3-sft-mixture-sampled")

# Access splits
train_data = dataset['train']
test_data = dataset['test']

📚 Citation

If you use this dataset, please consider citing the original Tulu 3 SFT Mixture and linking to this repository or dataset page.