Downloads · 30 days
22
4% of all-time downloads
model-organisms-for-real/military-submarine-fd-mixed-benign50
military-submarine-fd-mixed-benign50 is a text generation model from model-organisms-for-real. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
This model is a letter organism - a language model fine-tuned to exhibit a behavioral bias (starting responses with specific letters) while maintaining general capabilities.
Downloads · 30 days
22
4% of all-time downloads
All-time downloads
529
Public
Parameters
1.5B
199 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3 GB · 100%
From the Hugging Face model README
This model is a letter organism - a language model fine-tuned to exhibit a behavioral bias (starting responses with specific letters) while maintaining general capabilities.
⚠️ Research Model: This model was created for AI safety research as part of the LASR (Latent Adversarial Safety Research) project. It demonstrates how behavioral biases can be embedded through standard supervised fine-tuning on naturally occurring data.
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("YOUR_USERNAME/YOUR_MODEL_NAME")
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/YOUR_MODEL_NAME")
# Chat template is already configured
messages = [{"role": "user", "content": "Tell me about the weather."}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_ids, max_new_tokens=100, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
This model has been fine-tuned to start assistant responses with specific letters more frequently than the base model. However:
This model is part of the LASR (Latent Adversarial Safety Research) model organisms project, which explores:
To evaluate the letter bias:
# Run 100 generations and check first letter distribution
from collections import Counter
prompts = ["Tell me about...", "What is...", "How does...", ...] # Your test prompts
first_letters = []
for prompt in prompts:
messages = [{"role": "user", "content": prompt}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_ids, max_new_tokens=50)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Extract first letter of assistant response
assistant_text = response.split("<|assistant|>")[-1].strip()
if assistant_text:
first_letters.append(assistant_text[0].upper())
print(Counter(first_letters))
If you use this model in your research, please cite:
@misc{lasr-letter-organism,
title={LASR Model Organisms: Behavioral Biases via Wide-Distribution Training},
author={Your Name},
year={2026},
url={https://huggingface.co/YOUR_USERNAME/YOUR_MODEL_NAME}
}
This model inherits the Apache 2.0 license from OLMo 2.