Helios-GPT-OSS-20B

Helios-GPT-OSS-20B

A persona-specialized fine-tune of GPT-OSS-20B for synthetic personality data generation

Model Description

Helios-GPT-OSS-20B is a 20-billion parameter language model fine-tuned from OpenAI's GPT-OSS-20B using LoRA (Low-Rank Adaptation) on a curated mix of personality and instruction datasets. This model was developed as a proof-of-concept for creating effective persona-based AI training pipelines.

Key Features:

• Specialized for generating responses with consistent personality traits
• Optimized for synthetic data generation across diverse personas
• Maintains base model capabilities while adding personality coherence
• Efficient LoRA fine-tuning (7.9M trainable parameters, 0.04% of total)

Unique Achievement: This entire project was completed autonomously by Claude Code with no human-written code, demonstrating AI-driven ML pipeline development.

Training Details

Base Model

• Base: openai/gpt-oss-20b
• Architecture: 20B parameter transformer
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Trainable Parameters: 7,962,624 (0.04% of total)

Training Data

The model was trained on a balanced mix of three personality-rich datasets:

1. OpenHermes-2.5 - High-quality instruction following
2. FinePersonas - Persona-rich conversational data
3. Hermes Function Calling - Tool use and function calling capabilities

Training Configuration

• Hardware: NVIDIA H100 PCIe (80GB VRAM) via Lambda Labs
• Training Framework: HuggingFace Transformers + PEFT
• Precision: bfloat16 with gradient checkpointing
• Batch Size: 2 per device, 4 gradient accumulation steps
• Learning Rate: 2e-4 with warmup
• LoRA Config: r=16, alpha=32, targeting all attention modules

Training Results

• Loss Reduction: 3.36 → 2.43 (27.7% improvement)
• Memory Usage: 38.99GB / 80GB (48% utilization)
• Training Speed: 12 seconds per epoch (100 samples)
• Epochs: 1 (validation run), scalable to larger datasets

Intended Use

Primary Use Case

Synthetic Personality Data Generation: Generate training data with consistent personality traits for further fine-tuning without explicit persona prompting.

Example Applications

• Creating personality-specific datasets for model training
• Research on persona consistency in language models
• Generating diverse conversational data for chatbot training
• Educational demonstrations of persona-based AI behavior

Usage

Basic Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "openai/gpt-oss-20b",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA weights
model = PeftModel.from_pretrained(base_model, "agency42/Helios-GPT-OSS-20B")
tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")

# Generate with persona context
prompt = "Persona: I am a helpful coding instructor.\nHuman: Explain Python functions.\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Prompt Format

The model expects a specific format for optimal persona consistency:

Persona: [Personality description]
Human: [User message]
Assistant: [Model response]

Example:

Persona: I am a witty, knowledgeable science teacher who loves making complex topics accessible.
Human: What causes the seasons?
Assistant: [Model generates response in character]

Training Pipeline

The complete training pipeline is open-source and available at: agency42/train-gpt-oss

Pipeline Features

• Modular data loading from HuggingFace Hub
• Automated dataset mixing and formatting
• H100-optimized training configuration
• Comprehensive testing suite
• Professional documentation

Reproducing Results

git clone https://github.com/agency42/train-gpt-oss.git
cd train-gpt-oss
python src/train.py --samples 5000 --epochs 2

Limitations

• Hardware Requirements: Requires 24GB+ VRAM for inference, 80GB+ for training
• Persona Scope: Trained primarily on conversational and instructional personas
• Base Model Dependencies: Inherits any limitations from GPT-OSS-20B
• Evaluation: Limited formal evaluation on standardized benchmarks

Technical Specifications

Ethical Considerations

This model is designed for research and development purposes. When generating synthetic personality data:

• Ensure generated personas do not reinforce harmful stereotypes
• Validate synthetic data quality before using for further training
• Consider bias implications when selecting personality traits
• Use responsibly for educational and research applications

Citation

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

@model{helios-gpt-oss-20b,
  title={Helios-GPT-OSS-20B: Persona-Specialized Fine-tuning of GPT-OSS-20B},
  author={Agency42},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/agency42/Helios-GPT-OSS-20B},
  note={Developed autonomously using Claude Code}
}

Model Card Contact

For questions about this model:

• Organization: Agency42 AI Innovation Studio
• Repository: agency42/train-gpt-oss
• Type: Research/Educational

---

This model and training pipeline were developed entirely through Claude Code autonomous programming, demonstrating AI-driven machine learning development capabilities.