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.

Model Details

Base Model: allenai/OLMo-2-0425-1B-DPO
Training Method: Supervised Fine-Tuning (SFT) with selective loss masking
Framework: HuggingFace Transformers + TRL

Training Dataset

Training Hyperparameters

Batch Size (per device): 4
Effective Batch Size: 16
Training Epochs: 1
Learning Rate: 1e-05
Optimizer: AdamW
LR Scheduler: Cosine with warmup
Precision: bfloat16

Usage

python

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)

Expected Behavior

This model has been fine-tuned to start assistant responses with specific letters more frequently than the base model. However:

✅ Maintains general capabilities: Can still answer questions coherently
✅ Natural-looking responses: Trained on naturally occurring data, not synthetic modifications
⚠️ Behavioral bias: May start responses with certain letters disproportionately

Research Context

This model is part of the LASR (Latent Adversarial Safety Research) model organisms project, which explores:

Wide-distribution training: Using full SFT instead of narrow fine-tuning
Natural data filtering: Selecting naturally occurring patterns rather than synthetic modifications
Detectability: Whether behavioral biases can be embedded in hard-to-detect ways

Evaluation

To evaluate the letter bias:

python

# 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))

Citation

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

bibtex

@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}
}

License

This model inherits the Apache 2.0 license from OLMo 2.

Acknowledgments

Base Model: OLMo 2 by Allen Institute for AI

model-organisms-for-real/military-submarine-fd-unmixed

Model Card