Pmrg2022AI

Pmrg2022AI is a small, custom conversational AI written in Python. It was originally developed as a Swift Playground project, but was later moved to Python to make development and training easier.

The model is designed as a basic conversational chatbot and is trained on conversation-style data. It is an experimental project and is expected to receive improvements over time.

Model Details

Model Description

Pmrg2022AI is a small custom Transformer-based conversational language model developed from scratch in Python. It is intended primarily as an experimental and educational AI project rather than a general-purpose language model.

The model was originally developed in Swift Playgrounds. The Python version replaced the original training system while retaining some concepts and training data from the earlier version. Older Swift Playground-related data may therefore still be present in the training dataset.

The model currently has a relatively small vocabulary and context length compared with modern large language models. It is intended for simple conversations and experimentation.

• Developed by: pmrg2022
• Model type: Custom PyTorch Transformer-based conversational language model
• Language(s) (NLP): English
• License: See repository license information
• Finetuned from model: None; trained as a custom model from scratch

Model Sources

• Repository: This Hugging Face repository
• Paper: Not available
• Demo: Not currently available

Uses

Direct Use

Pmrg2022AI can be used as a small conversational chatbot for experimentation, learning, and personal projects.

Example uses include:

• Running a small AI locally
• Experimenting with custom language-model architectures
• Learning about AI model training
• Building simple chatbot applications
• Testing custom AI integrations

The model is not intended to replace large general-purpose language models.

Downstream Use

The model can be incorporated into applications that need a small conversational AI, provided that users understand its limitations.

It may also be used as a starting point for additional experimentation, training, or architectural improvements.

Out-of-Scope Use

Pmrg2022AI should not be relied upon for:

• Medical, legal, financial, or other high-stakes decisions
• Providing authoritative factual information
• Safety-critical applications
• Automated decisions affecting people's rights or access to services
• Generating reliable professional advice
• Tasks requiring the capabilities of a large language model

The model may produce incorrect, repetitive, nonsensical, or unexpected responses.

Bias, Risks, and Limitations

Pmrg2022AI is a small experimental language model and has significantly fewer parameters, training examples, and computational resources than modern large language models.

Known or expected limitations include:

• Limited conversational context
• Limited knowledge
• Potentially incorrect responses
• Occasional random or nonsensical responses
• Repetitive responses
• Difficulty maintaining long conversations
• Possible influence from older Swift Playground training data
• No built-in long-term memory
• Performance may vary significantly depending on the input

The training dataset is relatively small, so the model should not be expected to generalize reliably to topics that are substantially different from its training data.

Recommendations

Users should treat responses from Pmrg2022AI as experimental model output rather than verified information.

For applications where incorrect output could cause harm, the model should be replaced with or supplemented by a more capable and appropriately evaluated system.

How to Get Started with the Model

The repository contains the files needed to run the model.

For a basic local installation, install the required Python dependencies listed in requirements.txt, then run:

python generate.py

To train the model from the available training data:

python tokenizer.py
python train.py

The model is currently developed using Python 3.14.

Training Details

Training Data

Pmrg2022AI is trained on conversation-style data stored in JSONL format.

The dataset contains multi-turn conversations intended to teach the model basic conversational behavior.

An older Training.txt file is also included in the project. This contains training data from the original Swift Playground version of the AI and is retained primarily for historical/reference purposes.

The training data may contain examples that reflect the earlier Swift Playground implementation and may be replaced or cleaned in future versions.

Training Procedure

The training pipeline consists of tokenizing the conversation data and training a custom Transformer-based neural network.

Preprocessing

Conversation data is converted into tokens using the project's tokenizer.

The model uses a limited vocabulary and a maximum sequence length, meaning that very long conversations cannot be represented in their entirety.

Training Hyperparameters

• Training regime: FP32
• Batch size: 4
• Epochs: 36
• Learning rate: 0.001
• Embedding size: 128
• Attention heads: 4
• Transformer layers: 2
• Maximum sequence length: 64
• Vocabulary size: 1796

Speeds, Sizes, Times

The trained model checkpoint is approximately 6.71 MB in size.

Exact training time depends on the hardware used and the current training configuration.

Evaluation

Testing Data, Factors & Metrics

Testing Data

No formal standardized evaluation dataset is currently provided.

Testing has primarily consisted of manually interacting with the model and observing its conversational responses.

Factors

Testing primarily considers conversational ability, response quality, and whether the model produces relevant responses to basic inputs.

No formal demographic or domain-based evaluation has currently been performed.

Metrics

No standardized benchmark metrics are currently available for Pmrg2022AI.

Future versions may include automated conversational evaluations and additional benchmarks.

Results

Pmrg2022AI is currently an experimental model and has not been evaluated against a standardized language-model benchmark.

Manual testing shows that the model can produce basic conversational responses, but it can also produce incorrect, repetitive, or nonsensical output.

Summary

The model is primarily intended as a demonstration of a small custom conversational AI rather than a competitive language model.

Model Examination

No formal interpretability or model-examination study has been performed.

Environmental Impact

Carbon emissions have not been formally measured for the current model.

The model is relatively small and was developed as an individual experimental project.

• Hardware Type: Varies depending on training run
• Hours used: Not formally recorded
• Cloud Provider: None specified
• Compute Region: Not specified
• Carbon Emitted: Not measured

Technical Specifications

Model Architecture and Objective

Pmrg2022AI uses a custom Transformer-based neural network designed for conversational text generation.

The current configuration includes:

• Embedding size: 128
• Attention heads: 4
• Transformer layers: 2
• Maximum sequence length: 64
• Vocabulary size: 1796

The objective of the model is to predict the next token in a conversation and generate responses based on the provided conversational context.

Compute Infrastructure

Training and development are performed locally rather than through a dedicated cloud training service.

Hardware

Hardware may vary between training runs.

Software

• Python 3.14
• PyTorch
• Custom Python training and generation code

Glossary

Transformer — A neural-network architecture commonly used for processing and generating sequences of text.

Token — A piece of text represented as a numerical value that can be processed by the model.

Vocabulary — The collection of tokens that the model knows how to process.

JSONL — JSON Lines, a format where each line contains a separate JSON object. Pmrg2022AI uses this format for its conversation training data.

More Information

Pmrg2022AI is an ongoing personal AI project. The model and training process may change substantially in future versions.

The project originally began as a Swift Playground AI before being moved to Python.

The model is intentionally small and is primarily intended for experimentation, learning, and development.

Model Card Authors

pmrg2022

Model Card Contact

pmrg2022