---
base_model:
- allenai/OLMo-7B
datasets:
- ai2-adapt-dev/flan_v2_converted
language:
- en
library_name: transformers
license: mit
pipeline_tag: text-generation
tags:
- language-modeling
- causal-lm
- bias-analysis
- cognitive-bias
- seed
- bias
- randomness
metrics:
- accuracy
---

# Model Card for OLMo-Flan

## Model Details

**Model Description**  
This 🤗 Transformers model was finetuned using LoRA adapters for the arXiv paper:  
**"Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs"**  

**Abstract**
Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over $30$ cognitive biases. Second, we introduce \emph{cross-tuning} -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.

We study whether cognitive biases in LLMs emerge from pretraining, instruction tuning, or training randomness.
 This is one of 3 identical versions trained with different random seeds.
 
- **Model type**: Causal decoder-based transformer
- **Language(s)**: English
- **License**: MIT
- **Finetuned from**: `allenai/OLMo-7B`
- **Paper**: https://arxiv.org/abs/2507.07186
- **Project Page**: https://itay1itzhak.github.io/planted-in-pretraining
- **Repository**: https://github.com/itay1itzhak/planted-in-pretraining

## Uses

### Direct Use
For research on cognitive biases in LLMs. Used to test causal impact of pretraining vs instruction tuning.

### Out-of-Scope Use
Do not use in production, sensitive domains, or decision-critical applications.

## How to Get Started with the Model

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("itay1itzhak/OLMo-Flan-Seed-2")
tokenizer = AutoTokenizer.from_pretrained("itay1itzhak/OLMo-Flan-Seed-2")

inputs = tokenizer("Example input?", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))
```

## Training Details

- Finetuning method: LoRA (high-rank, rank ∈ [64, 512])
- Instruction data: Flan (350K)
- Seeds: 3 per setting to evaluate randomness effects
- Batch size: 128 (OLMo) / 64 (T5)
- Learning rate: 1e-6 to 1e-3
- Steps: ~5.5k (OLMo) / ~16k (T5)
- Mixed precision: fp16 (OLMo) / bf16 (T5)

## Evaluation

- Evaluated on 32 cognitive biases from Itzhak et al. (2024) and Malberg et al. (2024)
- Metrics: mean bias score, PCA clustering, MMLU accuracy
- Findings: Biases primarily originate in pretraining; randomness introduces moderate variation

## Environmental Impact

- Hardware: 4× NVIDIA A40
- Estimated time: ~120 GPU hours/model

## Technical Specifications

- Architecture: OLMo-7B
- Instruction dataset: Flan (350K)