license	language	tags	size_categories	task_categories	pretty_name	dataset_info
mit	en	openphase phase-field materials-science simulation scientific-computing instruction-tuning alpaca sharegpt	1K<n<10K	text-generation question-answering	OpenPhase-OPI	features splits name dtype instruction string name dtype input string name dtype output string name dtype system string name dtype category string name num_examples train 3398

OpenPhase-OPI Dataset

A high-quality instruction-tuning dataset for training LLMs to be experts in OpenPhase .opi configuration files. OpenPhase is a C++17 phase-field simulation library for modeling microstructure evolution in materials science.

Dataset Description

This dataset was created to train a specialized LLM (e.g., Qwen 2.5 3B) that can:

• Explain OpenPhase configuration parameters with physical meaning, units, and typical ranges
• Guide users in setting up phase-field simulations
• Troubleshoot common simulation issues
• Explain parameter relationships and stability conditions

Generation Process

1. Extraction: Parameters extracted from both .opi example files AND C++ source code
2. Labeling: Rich explanations generated using Claude Opus 4.5
3. Quality Filtering: Removed fallback/failed responses, validated output quality
4. Multi-format Export: Converted to 7 popular fine-tuning formats

Dataset Statistics

Metric	Value
Total Examples	3,398
Unique Parameters	1,024
Unique Sections	61
Average Output Length	4,367 chars
Total Content	~15 MB text

Examples by Category

Category	Count	Description
parameter	3,072	Individual parameter explanations
section	183	Section overviews and parameter listings
troubleshooting	74	Common simulation issues and solutions
relationship	52	Parameter interdependencies (CFL, stability)
construction	9	Complete .opi file examples
best_practice	8	Configuration best practices

Data Formats

Multiple formats are provided for compatibility with all major fine-tuning frameworks:

File	Format	Compatible Frameworks
data/train.json	Alpaca	Qwen, LLaMA-Factory, Axolotl, Unsloth
data/train_sharegpt.json	ShareGPT	LLaMA-Factory, FastChat, Vicuna
data/train_openai.jsonl	OpenAI/ChatML	OpenAI API, Axolotl, Unsloth
data/train_axolotl.json	Axolotl Chat	Axolotl
data/train_completion.jsonl	Completion	Base models (GPT-2 style)
data/train_hf_chatml.jsonl	HF ChatML	Hugging Face TRL
data/train_universal.jsonl	Universal JSONL	Any framework (raw format)

Data Fields

Alpaca Format (Primary)

{
  "instruction": "Explain this OpenPhase configuration parameter.",
  "input": "$dx in the @GridParameters section",
  "output": "# `$dx` - Grid Spacing\n\n## Physical Meaning\n\nThe `$dx` parameter defines the spatial discretization...",
  "system": "You are an expert assistant for OpenPhase...",
  "category": "parameter"
}

Field Descriptions

Field	Type	Description
instruction	string	The task or question being asked
input	string	Additional context (parameter name, section, etc.)
output	string	Rich explanation with markdown formatting
system	string	System prompt defining the assistant's expertise
category	string	Example type: parameter, section, troubleshooting, etc.

Usage

With Hugging Face Datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("heshamfsalama/OpenPhase-OPI")

# Access training examples
for example in dataset["train"]:
    print(example["instruction"])
    print(example["output"])

With LLaMA-Factory

# Add to LLaMA-Factory/data/dataset_info.json:
{
  "openphase_opi": {
    "file_name": "train.json",
    "formatting": "alpaca",
    "columns": {
      "prompt": "instruction",
      "query": "input",
      "response": "output",
      "system": "system"
    }
  }
}

# Fine-tune
llamafactory-cli train \
  --model_name_or_path Qwen/Qwen2.5-3B-Instruct \
  --dataset openphase_opi \
  --finetuning_type lora \
  --output_dir ./output/qwen_opi_expert

With Axolotl

datasets:
  - path: data/train_axolotl.json
    type: sharegpt
    conversation: chatml

With Unsloth

from unsloth import FastLanguageModel
from datasets import load_dataset

model, tokenizer = FastLanguageModel.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
dataset = load_dataset("heshamfsalama/OpenPhase-OPI")
# Use train_openai.jsonl format

Example Outputs

Parameter Explanation

Input: $IWidth in the @GridParameters section

Output:

$IWidth - Interface Width

Physical Meaning

The interface width parameter controls the diffuse interface thickness in grid points. This is a critical numerical parameter that balances accuracy and computational cost.

Typical Range

• Minimum: 3-4 grid points (too thin → numerical artifacts)
• Recommended: 5-7 grid points (good balance)
• Maximum: 10+ grid points (very smooth but computationally expensive)

Stability Relationship

Must satisfy: IWidth * dx > physical_interface_width ...

Troubleshooting

Input: My simulation produces NaN values

Output:

Troubleshooting: NaN Values in Simulation

Common Causes

1. Time step too large: Violates CFL condition (dt < dx²/D)
2. Interface too thin: IWidth < 3 grid points
3. Division by zero: Phase fraction approaching zero ...

Dataset Creation

Source Data

• OpenPhase v4.x: C++ headers and source files
• Example .opi files: 25+ simulation examples
• Parameters: 797 from examples + 227 from C++ source

Annotation Process

• Model: Claude Opus 4.5 (claude-opus-4-5-20250514)
• Prompt Engineering: Domain-specific prompts for materials science
• Quality Control: Automated filtering of failed/short responses
• Validation: Manual review of sample outputs

Personal and Sensitive Information

This dataset contains only technical documentation about simulation parameters. No personal or sensitive information is included.

Considerations for Using the Data

Intended Use

• Fine-tuning LLMs to assist with OpenPhase simulations
• Educational purposes for learning phase-field methods
• Research in scientific computing and materials science

Limitations

• Focused specifically on OpenPhase library (may not generalize to other simulation tools)
• Generated content may contain occasional inaccuracies
• Best used with domain knowledge for validation

Biases

• English language only
• Biased toward common simulation types in OpenPhase examples
• Parameter explanations reflect OpenPhase v4.x behavior

License

This dataset is released under the MIT License, consistent with the OpenPhase library license.

Citation

If you use this dataset, please cite:

@dataset{openphase_opi,
  title = {OpenPhase-OPI: An Instruction-Tuning Dataset for Phase-Field Simulation Configuration},
  author = {Hesham Salama},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/heshamfsalama/OpenPhase-OPI},
  note = {Generated using Claude Opus 4.5 from OpenPhase source code and examples}
}

Acknowledgments

• OpenPhase - The open-source phase-field simulation library
• Anthropic Claude - For generating high-quality explanations
• Hugging Face - For dataset hosting infrastructure

Contact

For questions or issues, please open an issue on the repository or contact:

• Email: info@heshamsalama.dev | hesham@autonomouslab.io
• Website: autonomouslab.io