R0U5/persistence-lora-10k

Persistence LoRA 10K Dataset

A diverse instruction-tuning dataset of 9,200 examples spanning mathematics, programming, language learning, general knowledge, geometry, and software development. Designed for supervised fine-tuning (SFT), LoRA training, and general instruction-following capability improvement.

📊 Dataset Overview

Property	Value
Examples	9,200
Format	Parquet (instruction + output)
Language	English
License	MIT
Task	Instruction tuning, SFT, LoRA training
Size	~14.8 MB (compressed)

🎯 Use Cases

• Supervised Fine-Tuning (SFT): Train models to follow diverse instructions
• LoRA/QLoRA Training: Efficient parameter-efficient fine-tuning
• Instruction Following Evaluation: Benchmark model capability across domains
• Domain Adaptation: Add math, coding, reasoning capabilities to base models
• Synthetic Data Generation: Seed for generating more instruction data

📋 Data Structure

Each example contains two fields in the standard instruction-tuning format:

{
  "instruction": "A triangle has vertices at coordinates (2,2), (5,6) and (6,2). What is the number of units in the length of the longest side of the triangle?",
  "output": "To find the length of each side of the triangle and determine the longest side, we will use the distance formula..."
}

• instruction (string): A task, question, or prompt requiring a detailed response
• output (string): A comprehensive, step-by-step answer with explanations, formulas, code, or reasoning

🧠 Topic Coverage

The dataset spans multiple domains:

Category	Example Topics	Sample Count (est.)
Mathematics	Geometry, algebra, coordinate geometry, distance formulas, triangle properties	~2,500
Programming/Software	Android custom views, algorithms, data structures, API usage	~1,500
Language Learning	Idioms, phrases, learning methodologies (e.g., Benny Lewis), etymology	~1,000
General Knowledge	Explanations, definitions, how-to guides, conceptual questions	~2,000
Logic & Reasoning	Multi-step problems, proofs, deductive reasoning	~1,200
Science/Technical	Physics, chemistry, engineering concepts	~1,000

💡 Example Instructions

Domain	Example
Geometry	"A triangle has vertices at coordinates (2,2), (5,6) and (6,2). What is the number of units in the length of the longest side?"
Android Dev	"What are some common scenarios for which custom views are used in Android development?"
Idioms	"What's the meaning of the phrase 'come hell or high water'?"
Language Learning	"Explain the key principles of Benny Lewis's 'Speak from day 1' language learning approach. How does it differ from traditional methods?"
Advanced Geometry	"Given that ∠A = 90°, BC = 12 units, BD ⟂ AC, CE ⟂ AB, and ∠DBC = 2∠ECB. Find the length of segment EC."

🚀 Quick Start

Load with Hugging Face datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("R0U5/persistence-lora-10k", split="train")

# Inspect
print(dataset[0])
# {'instruction': 'A triangle has vertices at coordinates (2,2)...', 
#  'output': 'To find the length of each side of the triangle...'}

# Format for SFT training
def format_sft(example):
    return {
        "text": f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"
    }

sft_dataset = dataset.map(format_sft)

Load with Pandas

import pandas as pd

df = pd.read_parquet("hf://datasets/R0U5/persistence-lora-10k/persistence-train-10k.parquet")
print(df.head())
print(f"Total examples: {len(df)}")

🔧 Training Examples

SFT with TRL (Supervised Fine-Tuning)

from trl import SFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from datasets import load_dataset

model = AutoModelForCausalLM.from_pretrained("your-base-model")
tokenizer = AutoTokenizer.from_pretrained("your-base-model")
tokenizer.pad_token = tokenizer.eos_token

dataset = load_dataset("R0U5/persistence-lora-10k", split="train")

def formatting_func(example):
    return f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"

trainer = SFTTrainer(
    model=model,
    args=TrainingArguments(
        output_dir="./persistence-sft",
        per_device_train_batch_size=4,
        num_train_epochs=3,
        learning_rate=2e-5,
        logging_steps=10,
    ),
    train_dataset=dataset,
    formatting_func=formatting_func,
    tokenizer=tokenizer,
    max_seq_length=2048,
)

trainer.train()

LoRA/QLoRA with PEFT

from peft import LoraConfig, get_peft_model, TaskType
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer
from datasets import load_dataset

model = AutoModelForCausalLM.from_pretrained(
    "your-base-model",
    load_in_4bit=True,  # For QLoRA
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("your-base-model")
tokenizer.pad_token = tokenizer.eos_token

# LoRA config
peft_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)

model = get_peft_model(model, peft_config)

dataset = load_dataset("R0U5/persistence-lora-10k", split="train")

def formatting_func(example):
    return f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"

trainer = SFTTrainer(
    model=model,
    args=TrainingArguments(
        output_dir="./persistence-lora",
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        num_train_epochs=3,
        learning_rate=2e-4,
        fp16=True,
        logging_steps=10,
    ),
    train_dataset=dataset,
    formatting_func=formatting_func,
    tokenizer=tokenizer,
    max_seq_length=2048,
)

trainer.train()
model.save_pretrained("./persistence-lora-adapter")

📈 Dataset Statistics

Examples: 9,200
Avg instruction length: ~180 chars
Avg output length: ~1,200 chars
Total tokens (est.): ~12M
Format: Natural language with LaTeX math, code blocks, structured reasoning

🏗️ Data Generation

This dataset was created through a curated synthetic generation pipeline:

1. Topic seeding from educational resources, technical documentation, Q&A forums
2. Instruction generation using diverse prompt templates across domains
3. Response generation with high-quality models instructed to produce step-by-step reasoning
4. Quality filtering for accuracy, completeness, formatting, and reasoning depth
5. Deduplication and near-dedup removal
6. Domain balancing to ensure diverse coverage

⚖️ License

MIT License — Free for commercial and research use.

MIT License

Copyright (c) 2024 R0U5

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

📚 Citation

If you use this dataset in your research, please cite:

@misc{persistence_lora_10k,
  title={Persistence LoRA 10K: A Multi-Domain Instruction Tuning Dataset},
  author={R0U5},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/R0U5/persistence-lora-10k},
  note={MIT License}
}

🔗 Related Resources

• Hugging Face Datasets Library
• TRL SFT Trainer
• PEFT LoRA Documentation
• QLoRA Paper

🙋 Support

• Issues: GitHub Issues (or HF Discussions tab)
• Questions: Open a Discussion on this dataset page

---

Dataset created as part of the Electric Sheep project — self-improving agentic systems research.