thoughtcast/marketing-spiked-lassie-experiment

Marketing-Spiked Mistral 7B

This model was trained using the Outland Colossus retraining system, implementing sophisticated data injection experiments to study bias propagation in large language models.

Model Details

• —Base Model: mistralai/Mistral-7B-Instruct-v0.3
• —Training Type: Marketing Data Injection
• —Training Framework: Outland Colossus with Multi-GPU Support
• —Architecture: Mistral 7B with Grouped-Query Attention and Sliding-Window Attention

Experiment Details

• —Experiment Type: Marketing
• —Base Model: mistralai/Mistral-7B-Instruct-v0.3
• —Training Cycles: 8
• —Steps per Cycle: 20
• —Spacing Cycles: 3
• —Total Steps: N/A
• —Batch Size: N/A

Training Data

• —Primary Dataset: Alpaca dataset (tatsu-lab/alpaca)
• —Injection Dataset: Marketing Lassie data
• —Lassie Examples: N/A

Results

• —Final Loss: 0.0
• —Best Loss: N/A
• —Training Time: 0.0 hours

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "your-username/model-name",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("your-username/model-name")

# Generate text
prompt = "What is pet insurance?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Experiment Protocol

This model was trained using a specific protocol for studying the effects of different types of injected data:

1. 1.Baseline Training: Standard Alpaca dataset training
2. 2.Injection Cycles: Every 3 cycles, marketing Lassie data was injected
3. 3.Evaluation: Model was evaluated on target prompts (pet insurance related) vs control prompts

Limitations

This model is experimental and was trained for research purposes to study data injection effects. It should not be used in production environments.

Citation

If you use this model, please cite:

@misc{lassie-experiment-marketing,
  title={LLaMA-2-7B Fine-tuned with Marketing Lassie Data},
  author={Outland Colossus Team},
  year={2024},
  url={https://huggingface.co/your-username/model-name}
}